rypow commited on
Commit
a391b12
·
verified ·
1 Parent(s): 9b4e435

v0.2.1 initial release: 19.3M tetragon events {train:16,951,950, test: 2,349,965} timespans: {[train: 2026-04-05 15:40:57.657541+00:00 -> 2026-04-19 22:09:18.103543+00:00], [test: 2026-04-18 22:41:25.297890+00:00 -> 2026-04-20 02:01:27.614304+00:00]} labeled test events: 705,852 / 2,349,965 (30.0% of test split)

Browse files
Files changed (36) hide show
  1. LICENSE +36 -0
  2. README.md +101 -0
  3. RELEASE.json +157 -0
  4. SHA256SUMS +35 -0
  5. data/test-00000-of-00002.parquet +3 -0
  6. data/test-00001-of-00002.parquet +3 -0
  7. data/train-00000-of-00009.parquet +3 -0
  8. data/train-00001-of-00009.parquet +3 -0
  9. data/train-00002-of-00009.parquet +3 -0
  10. data/train-00003-of-00009.parquet +3 -0
  11. data/train-00004-of-00009.parquet +3 -0
  12. data/train-00005-of-00009.parquet +3 -0
  13. data/train-00006-of-00009.parquet +3 -0
  14. data/train-00007-of-00009.parquet +3 -0
  15. data/train-00008-of-00009.parquet +3 -0
  16. datacard.md +425 -0
  17. datasheet.md +413 -0
  18. dendroaspis.bib +13 -0
  19. labels.csv +62 -0
  20. reproduction/PARSER_VERSION +19 -0
  21. reproduction/REPRODUCE.md +216 -0
  22. reproduction/alias_map.json +333 -0
  23. reproduction/allowlist_basenames.txt +308 -0
  24. reproduction/art-campaign/ARTcampaign.md +383 -0
  25. reproduction/art-campaign/labels.csv +62 -0
  26. reproduction/feature_definitions.json +821 -0
  27. reproduction/requirements.txt +24 -0
  28. reproduction/tetragon-configs/identity_map.yaml +81 -0
  29. reproduction/tetragon-configs/tetragon/attribute-monitor.yaml +41 -0
  30. reproduction/tetragon-configs/tetragon/file-monitor.yaml +46 -0
  31. reproduction/tetragon-configs/tetragon/injection-monitor.yaml +82 -0
  32. reproduction/tetragon-configs/tetragon/logrotate.tetragon +25 -0
  33. reproduction/tetragon-configs/tetragon/memory-monitor.yaml +39 -0
  34. reproduction/tetragon-configs/tetragon/network-monitor.yaml +80 -0
  35. reproduction/tetragon-configs/tetragon/privilege-monitor.yaml +29 -0
  36. reproduction/tetragon-configs/tetragon_parser.yaml +43 -0
LICENSE ADDED
@@ -0,0 +1,36 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Attribution 4.0 International (CC BY 4.0)
2
+
3
+ Copyright (c) 2026 Ryan William Powers
4
+
5
+ This dataset is licensed under the Creative Commons Attribution 4.0
6
+ International License. You are free to:
7
+
8
+ - Share — copy and redistribute the material in any medium or format
9
+ - Adapt — remix, transform, and build upon the material
10
+ for any purpose, even commercially.
11
+
12
+ The licensor cannot revoke these freedoms as long as you follow the license
13
+ terms. Under the following terms:
14
+
15
+ Attribution — You must give appropriate credit, provide a link to the
16
+ license, and indicate if changes were made. You may do so in any
17
+ reasonable manner, but not in any way that suggests the licensor
18
+ endorses you or your use.
19
+
20
+ No additional restrictions — You may not apply legal terms or
21
+ technological measures that legally restrict others from doing anything
22
+ the license permits.
23
+
24
+ Notices:
25
+
26
+ You do not have to comply with the license for elements of the material
27
+ in the public domain or where your use is permitted by an applicable
28
+ exception or limitation.
29
+
30
+ No warranties are given. The license may not give you all of the
31
+ permissions necessary for your intended use. For example, other rights
32
+ such as publicity, privacy, or moral rights may limit how you use the
33
+ material.
34
+
35
+ For the full legal text, see:
36
+ https://creativecommons.org/licenses/by/4.0/legalcode
README.md ADDED
@@ -0,0 +1,101 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # v0.2 Tetragon-native HIDS corpus
2
+
3
+ **22 technique IDs (21 ATT&CK + 1 capstone) · 19.3M events · 6-day controlled collection window**
4
+
5
+ A behavioral-event corpus collected from a single Linux LXC container under a
6
+ controlled Atomic Red Team (ART) campaign, parsed and feature-engineered with
7
+ a Tetragon-native pipeline. Released for host-anomaly-detection research,
8
+ sequence-modeling baselines, and ATT&CK-technique-detectability studies.
9
+
10
+ This is the **v0.2.1 release** — the same campaign and Tetragon collection as
11
+ v0.2.0, re-parsed on 2026-04-29 with five correctness fixes (procfs-walk
12
+ tag-not-drop, per-process `delta_t` keying, train→test lineage seeding,
13
+ parallel-parser DuckDB sort, NLL alignment) and the 3a′ encoder revisions
14
+ (hash right-sizing, cwd depth 2→4, four promoted categoricals, dropped
15
+ `f_in_init_tree`). See `reproduction/PARSER_VERSION` for the full changelog.
16
+
17
+ ## Contents
18
+
19
+ | Path | Purpose |
20
+ |-------------------------------|---------|
21
+ | `data/train-XXXXX-of-00009.parquet` | 9 shards · 16,951,950 rows · benign-only by construction |
22
+ | `data/test-XXXXX-of-00002.parquet` | 2 shards · 2,349,965 rows · 705,852 attack events (~30 %) |
23
+ | `labels.csv` | 61 ART intervals across 22 technique IDs |
24
+ | `datacard.md` | Hugging Face datacard — full feature inventory + biases + headline numbers |
25
+ | `datasheet.md` | Gebru-format datasheet — provenance + collection process |
26
+ | `LICENSE` | CC BY 4.0 |
27
+ | `dendroaspis.bib` | BibTeX citation block |
28
+ | `SHA256SUMS` | Integrity manifest |
29
+ | `ARCHIVED.md` | Pointer to prior bundle revisions kept under `archive/` |
30
+ | `reproduction/` | Tetragon configs + parser pin + REPRODUCE.md + feature_definitions.json + alias_map.json |
31
+
32
+ ## Quick start
33
+
34
+ ```python
35
+ from datasets import load_dataset
36
+ import pandas as pd
37
+
38
+ ds = load_dataset("<user>/dendroaspis-tetragon-hids", streaming=True)
39
+ event = next(iter(ds["train"]))
40
+ print(event["f_event_type"], event["f_action_family"], event["token"])
41
+
42
+ labels = pd.read_csv("labels.csv", parse_dates=["start_ts", "end_ts"])
43
+ labels["start_ts"] = labels["start_ts"].dt.tz_convert("UTC")
44
+ labels["end_ts"] = labels["end_ts"].dt.tz_convert("UTC")
45
+ print(f"{len(labels)} intervals · {labels['technique_id'].nunique()} techniques")
46
+ ```
47
+
48
+ To interpret the integer-coded categorical columns, load
49
+ `reproduction/feature_definitions.json` — every code (e.g. `f_action_family=6`)
50
+ maps to a label and a one-line semantic description.
51
+
52
+ ## Splits at a glance
53
+
54
+ | Split | Rows | Collection window (UTC)† | Attack rate |
55
+ |-------|------------|------------------------------------------------|-------------|
56
+ | train | 16,951,950 | 2026-04-14 01:41 → 2026-04-19 22:09 | 0.0 % |
57
+ | test | 2,349,965 | 2026-04-19 21:46 → 2026-04-20 02:01 | 30.0 % |
58
+
59
+ All ART attack intervals fall in the **test** split. Train is benign-only;
60
+ unsupervised / self-supervised methods are the natural fit.
61
+
62
+ **† Pre-window event timestamps.** ≈295 events (267 train + 28 test, 0.0015 %
63
+ of the corpus) carry `event_time` values that predate the stated collection
64
+ window. All have `f_is_procfs_walk = 1`: they are parser-synthesized
65
+ `process_exec` records for daemons that were already running when Tetragon
66
+ collection started, using the daemon's original `process_start_time` as the
67
+ event timestamp. They preserve lineage roots for downstream events. Filter
68
+ `WHERE f_is_procfs_walk = 0` to exclude them from time-window analysis.
69
+
70
+ ## Headline numbers
71
+
72
+ See `datacard.md` for the canonical Mamba NLL / MEM-FA AUROCs and the
73
+ strongest classical-baseline comparison. The `datacard` is the citation
74
+ target for any specific number; this README intentionally omits a single
75
+ headline figure to remain stable across eval refinements.
76
+
77
+ ## Citation
78
+
79
+ ```bibtex
80
+ @dataset{dendroaspis_v0_2_2026,
81
+ title = {v0.2 Tetragon-native HIDS corpus with 21 ATT\&CK techniques},
82
+ author = {Powers, Ryan William},
83
+ year = {2026},
84
+ version = {0.2.1}
85
+ }
86
+ ```
87
+
88
+ ## Pointers
89
+
90
+ - **Full feature documentation:** [datacard.md](./datacard.md)
91
+ - **Provenance + collection process:** [datasheet.md](./datasheet.md)
92
+ - **Reproducing collection:** [reproduction/REPRODUCE.md](./reproduction/REPRODUCE.md)
93
+ - **Integer-code → label tables:** [reproduction/feature_definitions.json](./reproduction/feature_definitions.json)
94
+ - **Anonymizer alias map (basenames):** [reproduction/alias_map.json](./reproduction/alias_map.json)
95
+
96
+ ## Intended use
97
+
98
+ Defensive research only. The attack telemetry is real ATT&CK technique
99
+ execution but generated in an isolated lab LXC container; the corpus is
100
+ **not** an attack template or payload library. See `datacard.md` "Intended
101
+ use + dual-use disclaimer" for the full statement.
RELEASE.json ADDED
@@ -0,0 +1,157 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": "0.2.1",
3
+ "build_timestamp": "2026-04-30T06:57:26Z",
4
+ "huggingface_repo": null,
5
+ "huggingface_revision": null,
6
+ "anonymizer_version": "0.2.1",
7
+ "parser_sha": "03e2204",
8
+ "release_parser_sha": "03e2204",
9
+ "row_counts": {
10
+ "train": 16951950,
11
+ "test": 2349965
12
+ },
13
+ "file_count": 35,
14
+ "total_bytes": 171517332,
15
+ "files": {
16
+ "ARCHIVED.md": {
17
+ "sha256": "b31f07a88de1477a75305afd46249323715d6d4574fa6570f201205c1e8b2409",
18
+ "size_bytes": 2958
19
+ },
20
+ "LICENSE": {
21
+ "sha256": "fa42b7b2d283cc859edccfeb9527996b575dc4043a2f007349a4b410205151e5",
22
+ "size_bytes": 1372
23
+ },
24
+ "README.md": {
25
+ "sha256": "61fc64236c5cca59d757562e75749cb7e66f2452e80d0592fc2b5b2f1f1ef2f6",
26
+ "size_bytes": 4871
27
+ },
28
+ "data/test-00000-of-00002.parquet": {
29
+ "sha256": "ba55902d9c0bbb9bcc28c8a3ee4fec6cc8acea797f0a38d642e3617e2dbc5c3a",
30
+ "size_bytes": 16847273
31
+ },
32
+ "data/test-00001-of-00002.parquet": {
33
+ "sha256": "e32756b18e4e6cdb35c9bd2369d934baeaee195ff78d80d240b5a174b031ab6b",
34
+ "size_bytes": 3659263
35
+ },
36
+ "data/train-00000-of-00009.parquet": {
37
+ "sha256": "6de5d6c6db629855c675d44622e75c8dacd942ced3fe3e6ebbce5eb7f4a5f891",
38
+ "size_bytes": 18967327
39
+ },
40
+ "data/train-00001-of-00009.parquet": {
41
+ "sha256": "0584d50bcfe191768f779994df453b7b3f063083a25c0751259cb2e8960a4ae8",
42
+ "size_bytes": 19780882
43
+ },
44
+ "data/train-00002-of-00009.parquet": {
45
+ "sha256": "ad290a091d5c752b1b224291befa67f30af74ccb1cd60ff5fa2e477dd975b21e",
46
+ "size_bytes": 17405683
47
+ },
48
+ "data/train-00003-of-00009.parquet": {
49
+ "sha256": "29f1bd4cc37772651a0fbf18eb02014f32b242c737640151e049666ec30f4330",
50
+ "size_bytes": 18115162
51
+ },
52
+ "data/train-00004-of-00009.parquet": {
53
+ "sha256": "54262554c7dee45fac124b8e06984a49b024e8eee0c38ffcc97f85b9086908be",
54
+ "size_bytes": 18282185
55
+ },
56
+ "data/train-00005-of-00009.parquet": {
57
+ "sha256": "9d2e176d0a929ce24e908a597f2e9c05fd20187842625de7979094aebfa2d8b5",
58
+ "size_bytes": 18918206
59
+ },
60
+ "data/train-00006-of-00009.parquet": {
61
+ "sha256": "da9ec971eeb85d5fee9fac52b27c0c41c3361a74196947ce42653b852d61b72d",
62
+ "size_bytes": 15642911
63
+ },
64
+ "data/train-00007-of-00009.parquet": {
65
+ "sha256": "6c88c52388ced41c9bfc23f39788d66390c4f8a1deb2217b185a59fb613d910b",
66
+ "size_bytes": 14969731
67
+ },
68
+ "data/train-00008-of-00009.parquet": {
69
+ "sha256": "ff4809c170ee636ae2a1199726b0f2343ff1ada908b7eac4042947e138ac7648",
70
+ "size_bytes": 8768242
71
+ },
72
+ "datacard.md": {
73
+ "sha256": "6b6cc845ec07a2b817db949bdc4ebcc42418778235740479ec2129f764785a20",
74
+ "size_bytes": 21842
75
+ },
76
+ "datasheet.md": {
77
+ "sha256": "85192fda0fae6c4c6b57dc1a6ac9eef7821ad2d7f1f10a735a99242480867f72",
78
+ "size_bytes": 17033
79
+ },
80
+ "dendroaspis.bib": {
81
+ "sha256": "9bbec25b813900edf746e2f6b053965d0d665dd70b6884e3b89c7c851e38681d",
82
+ "size_bytes": 541
83
+ },
84
+ "labels.csv": {
85
+ "sha256": "7b4f674b9a34aea092982c53cb2621c53b827058471c2b16ac1c7abe61668e03",
86
+ "size_bytes": 8286
87
+ },
88
+ "reproduction/PARSER_VERSION": {
89
+ "sha256": "bc7ae49d8398050b61f623afbc300bdf7f402ee5d823dc5c4dc739967cf51d6b",
90
+ "size_bytes": 1379
91
+ },
92
+ "reproduction/REPRODUCE.md": {
93
+ "sha256": "42b53c98042ebb0f7ece5214464ec559254cb43a6911471064be78fda06ee0ae",
94
+ "size_bytes": 8868
95
+ },
96
+ "reproduction/alias_map.json": {
97
+ "sha256": "398c52bfa3191c03c8d69f3c2ffb91bf17b672bbd9be9e70904ffd6e315f0ad6",
98
+ "size_bytes": 10026
99
+ },
100
+ "reproduction/allowlist_basenames.txt": {
101
+ "sha256": "9919392e9d46de602de1ac462eaf531ad35a9a9eded777ef7c37320f4f75ea2a",
102
+ "size_bytes": 3333
103
+ },
104
+ "reproduction/art-campaign/ARTcampaign.md": {
105
+ "sha256": "5d3e4ded07e8666d3a8c3ceecae6149309cf38520b25caa984d6fd8421c66e55",
106
+ "size_bytes": 28300
107
+ },
108
+ "reproduction/art-campaign/labels.csv": {
109
+ "sha256": "7b4f674b9a34aea092982c53cb2621c53b827058471c2b16ac1c7abe61668e03",
110
+ "size_bytes": 8286
111
+ },
112
+ "reproduction/feature_definitions.json": {
113
+ "sha256": "41160c9ab6e74ab035a281ba63aab593d1827883f13a131e4a4826afbb48c3ea",
114
+ "size_bytes": 24944
115
+ },
116
+ "reproduction/requirements.txt": {
117
+ "sha256": "90c49808a91028355aa630f0205bbd3f7fdfa3000e0c7a6da09e18c4e05a8ac6",
118
+ "size_bytes": 369
119
+ },
120
+ "reproduction/tetragon-configs/identity_map.yaml": {
121
+ "sha256": "a509ece1d28792404077dad0b93436c78689eb520b851784cf60936db879aadb",
122
+ "size_bytes": 3714
123
+ },
124
+ "reproduction/tetragon-configs/tetragon/attribute-monitor.yaml": {
125
+ "sha256": "892d549f7b6d26c08dfaf38fbc054b139822f0115aba4a58d0b25691b2c08c1c",
126
+ "size_bytes": 1725
127
+ },
128
+ "reproduction/tetragon-configs/tetragon/file-monitor.yaml": {
129
+ "sha256": "21405118c10c7700fe04414978655432e20b41a91daf429dfab35c63861b9d2b",
130
+ "size_bytes": 1467
131
+ },
132
+ "reproduction/tetragon-configs/tetragon/injection-monitor.yaml": {
133
+ "sha256": "4ecd684a2fad114335fbee836d3cdb74a4d4b7fc3654691d364111f1668a9b78",
134
+ "size_bytes": 3161
135
+ },
136
+ "reproduction/tetragon-configs/tetragon/logrotate.tetragon": {
137
+ "sha256": "20af1e30e99cabf327547ec1cf1972482547aa27191140a4eaab7b5b2c857553",
138
+ "size_bytes": 798
139
+ },
140
+ "reproduction/tetragon-configs/tetragon/memory-monitor.yaml": {
141
+ "sha256": "e60fb71b286a67ee3ecb065cd9243c0d04cc2cf6164ab5a83607f68bf6ceb8bf",
142
+ "size_bytes": 1591
143
+ },
144
+ "reproduction/tetragon-configs/tetragon/network-monitor.yaml": {
145
+ "sha256": "02cf3c59381c6eec7fe477f81a9462ce6bd6403850ce69876df4e6fe065299d4",
146
+ "size_bytes": 2571
147
+ },
148
+ "reproduction/tetragon-configs/tetragon/privilege-monitor.yaml": {
149
+ "sha256": "a0dece74ce8709f0135cef327a85f6fe8b40a8d48270fd636ce5006962b6fd75",
150
+ "size_bytes": 996
151
+ },
152
+ "reproduction/tetragon-configs/tetragon_parser.yaml": {
153
+ "sha256": "e65f6c8ad21ab3ad55bf696a9432fb3cad761b839991b84cb7faa89889a588fd",
154
+ "size_bytes": 2036
155
+ }
156
+ }
157
+ }
SHA256SUMS ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ b31f07a88de1477a75305afd46249323715d6d4574fa6570f201205c1e8b2409 ARCHIVED.md
2
+ fa42b7b2d283cc859edccfeb9527996b575dc4043a2f007349a4b410205151e5 LICENSE
3
+ 61fc64236c5cca59d757562e75749cb7e66f2452e80d0592fc2b5b2f1f1ef2f6 README.md
4
+ ba55902d9c0bbb9bcc28c8a3ee4fec6cc8acea797f0a38d642e3617e2dbc5c3a data/test-00000-of-00002.parquet
5
+ e32756b18e4e6cdb35c9bd2369d934baeaee195ff78d80d240b5a174b031ab6b data/test-00001-of-00002.parquet
6
+ 6de5d6c6db629855c675d44622e75c8dacd942ced3fe3e6ebbce5eb7f4a5f891 data/train-00000-of-00009.parquet
7
+ 0584d50bcfe191768f779994df453b7b3f063083a25c0751259cb2e8960a4ae8 data/train-00001-of-00009.parquet
8
+ ad290a091d5c752b1b224291befa67f30af74ccb1cd60ff5fa2e477dd975b21e data/train-00002-of-00009.parquet
9
+ 29f1bd4cc37772651a0fbf18eb02014f32b242c737640151e049666ec30f4330 data/train-00003-of-00009.parquet
10
+ 54262554c7dee45fac124b8e06984a49b024e8eee0c38ffcc97f85b9086908be data/train-00004-of-00009.parquet
11
+ 9d2e176d0a929ce24e908a597f2e9c05fd20187842625de7979094aebfa2d8b5 data/train-00005-of-00009.parquet
12
+ da9ec971eeb85d5fee9fac52b27c0c41c3361a74196947ce42653b852d61b72d data/train-00006-of-00009.parquet
13
+ 6c88c52388ced41c9bfc23f39788d66390c4f8a1deb2217b185a59fb613d910b data/train-00007-of-00009.parquet
14
+ ff4809c170ee636ae2a1199726b0f2343ff1ada908b7eac4042947e138ac7648 data/train-00008-of-00009.parquet
15
+ 6b6cc845ec07a2b817db949bdc4ebcc42418778235740479ec2129f764785a20 datacard.md
16
+ 85192fda0fae6c4c6b57dc1a6ac9eef7821ad2d7f1f10a735a99242480867f72 datasheet.md
17
+ 9bbec25b813900edf746e2f6b053965d0d665dd70b6884e3b89c7c851e38681d dendroaspis.bib
18
+ 7b4f674b9a34aea092982c53cb2621c53b827058471c2b16ac1c7abe61668e03 labels.csv
19
+ bc7ae49d8398050b61f623afbc300bdf7f402ee5d823dc5c4dc739967cf51d6b reproduction/PARSER_VERSION
20
+ 42b53c98042ebb0f7ece5214464ec559254cb43a6911471064be78fda06ee0ae reproduction/REPRODUCE.md
21
+ 398c52bfa3191c03c8d69f3c2ffb91bf17b672bbd9be9e70904ffd6e315f0ad6 reproduction/alias_map.json
22
+ 9919392e9d46de602de1ac462eaf531ad35a9a9eded777ef7c37320f4f75ea2a reproduction/allowlist_basenames.txt
23
+ 5d3e4ded07e8666d3a8c3ceecae6149309cf38520b25caa984d6fd8421c66e55 reproduction/art-campaign/ARTcampaign.md
24
+ 7b4f674b9a34aea092982c53cb2621c53b827058471c2b16ac1c7abe61668e03 reproduction/art-campaign/labels.csv
25
+ 41160c9ab6e74ab035a281ba63aab593d1827883f13a131e4a4826afbb48c3ea reproduction/feature_definitions.json
26
+ 90c49808a91028355aa630f0205bbd3f7fdfa3000e0c7a6da09e18c4e05a8ac6 reproduction/requirements.txt
27
+ a509ece1d28792404077dad0b93436c78689eb520b851784cf60936db879aadb reproduction/tetragon-configs/identity_map.yaml
28
+ 892d549f7b6d26c08dfaf38fbc054b139822f0115aba4a58d0b25691b2c08c1c reproduction/tetragon-configs/tetragon/attribute-monitor.yaml
29
+ 21405118c10c7700fe04414978655432e20b41a91daf429dfab35c63861b9d2b reproduction/tetragon-configs/tetragon/file-monitor.yaml
30
+ 4ecd684a2fad114335fbee836d3cdb74a4d4b7fc3654691d364111f1668a9b78 reproduction/tetragon-configs/tetragon/injection-monitor.yaml
31
+ 20af1e30e99cabf327547ec1cf1972482547aa27191140a4eaab7b5b2c857553 reproduction/tetragon-configs/tetragon/logrotate.tetragon
32
+ e60fb71b286a67ee3ecb065cd9243c0d04cc2cf6164ab5a83607f68bf6ceb8bf reproduction/tetragon-configs/tetragon/memory-monitor.yaml
33
+ 02cf3c59381c6eec7fe477f81a9462ce6bd6403850ce69876df4e6fe065299d4 reproduction/tetragon-configs/tetragon/network-monitor.yaml
34
+ a0dece74ce8709f0135cef327a85f6fe8b40a8d48270fd636ce5006962b6fd75 reproduction/tetragon-configs/tetragon/privilege-monitor.yaml
35
+ e65f6c8ad21ab3ad55bf696a9432fb3cad761b839991b84cb7faa89889a588fd reproduction/tetragon-configs/tetragon_parser.yaml
data/test-00000-of-00002.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ba55902d9c0bbb9bcc28c8a3ee4fec6cc8acea797f0a38d642e3617e2dbc5c3a
3
+ size 16847273
data/test-00001-of-00002.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e32756b18e4e6cdb35c9bd2369d934baeaee195ff78d80d240b5a174b031ab6b
3
+ size 3659263
data/train-00000-of-00009.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6de5d6c6db629855c675d44622e75c8dacd942ced3fe3e6ebbce5eb7f4a5f891
3
+ size 18967327
data/train-00001-of-00009.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0584d50bcfe191768f779994df453b7b3f063083a25c0751259cb2e8960a4ae8
3
+ size 19780882
data/train-00002-of-00009.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ad290a091d5c752b1b224291befa67f30af74ccb1cd60ff5fa2e477dd975b21e
3
+ size 17405683
data/train-00003-of-00009.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:29f1bd4cc37772651a0fbf18eb02014f32b242c737640151e049666ec30f4330
3
+ size 18115162
data/train-00004-of-00009.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:54262554c7dee45fac124b8e06984a49b024e8eee0c38ffcc97f85b9086908be
3
+ size 18282185
data/train-00005-of-00009.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9d2e176d0a929ce24e908a597f2e9c05fd20187842625de7979094aebfa2d8b5
3
+ size 18918206
data/train-00006-of-00009.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:da9ec971eeb85d5fee9fac52b27c0c41c3361a74196947ce42653b852d61b72d
3
+ size 15642911
data/train-00007-of-00009.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6c88c52388ced41c9bfc23f39788d66390c4f8a1deb2217b185a59fb613d910b
3
+ size 14969731
data/train-00008-of-00009.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ff4809c170ee636ae2a1199726b0f2343ff1ada908b7eac4042947e138ac7648
3
+ size 8768242
datacard.md ADDED
@@ -0,0 +1,425 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-4.0
3
+ task_categories:
4
+ - tabular-classification
5
+ - other
6
+ tags:
7
+ - host-intrusion-detection
8
+ - tetragon
9
+ - ebpf
10
+ - sequence-modeling
11
+ - att-and-ck
12
+ - anomaly-detection
13
+ size_categories:
14
+ - 10M<n<100M
15
+ language: []
16
+ pretty_name: "v0.2 Tetragon-native HIDS corpus (22 technique IDs)"
17
+ ---
18
+
19
+ # v0.2 — Tetragon-native HIDS corpus with 22 technique IDs (21 ATT&CK + 1 capstone)
20
+
21
+ A behavioral-event corpus collected from a single Linux host under a
22
+ controlled Atomic Red Team (ART) campaign. ~19.3 million events (process /
23
+ file / network / credential / memory) emitted by Tetragon eBPF probes inside
24
+ an LXC container, parsed and feature-engineered into a model-ready columnar
25
+ parquet. 21 ATT&CK techniques across 8 tactics plus a one-trial supply-chain
26
+ capstone are time-window-labeled in `labels.csv`. Released alongside the
27
+ **v0.2 paper** (course milestone) that uses this corpus to compare an
28
+ autoregressive Mamba sequence model against masked-event-modeling and
29
+ classical baselines.
30
+
31
+ This is the **v0.2.1 release** — same Tetragon collection as v0.2.0, re-parsed
32
+ on 2026-04-29 with five correctness fixes plus the 3a′ encoder revisions
33
+ (see *Schema evolution* below and `reproduction/PARSER_VERSION`).
34
+
35
+ ## Headline numbers (Run 2, post-rebuild)
36
+
37
+ | Scorer | Test AUROC | Notes |
38
+ |-------------------------------------|------------|-------|
39
+ | Mamba MEM-FA (field-aware) | **0.7513** | Best Mamba result; mask-modeling with per-field categorical CE heads. |
40
+ | Mamba NLL (autoregressive) | 0.7411 | Next-event NLL with the alignment fix. |
41
+ | IsolationForest (best classical) | 0.6658 | Trained on the same encoded feature columns. |
42
+ | XGBoost | 0.5000 | Degenerate on this corpus — train has no positives. |
43
+ | n-gram (token-level) | 0.5672 | Per-window aggregate; inverts on T1083 + T1059.004 (Predictability Paradox). |
44
+
45
+ Mamba's sequence stack adds ≈0.085 AUROC over the strongest classical
46
+ (IsolationForest). Per-technique breakdowns and the per-window inversion on
47
+ unsupervised classicals are reported in the v0.2 paper §5; raw evaluation
48
+ JSONs live at `mamba-edge/artifacts/v0.2/` in the project repository.
49
+
50
+ ---
51
+
52
+ ## Dataset summary
53
+
54
+ - **Source:** Tetragon 1.6.1 eBPF telemetry on a single LXC container,
55
+ 6-day collection window 2026-04-14 → 2026-04-20.
56
+ - **Population:** all process / file / network / credential / memory events
57
+ attributable to the target container's PID-namespace inum across the
58
+ collection window (parser filters host-namespace events at parse time).
59
+ - **Splits:** chronological — train (2026-04-14T01:41Z → 2026-04-19T22:09Z) /
60
+ test (2026-04-19T21:46Z → 2026-04-20T02:01Z). All ART attack intervals
61
+ fall in the test split; train is benign-only by construction.
62
+ - **Format:** Parquet shards under `data/`, suitable for the Hugging Face
63
+ `datasets` library streaming loader. Labels are a separate `labels.csv`.
64
+ - **License:** CC BY 4.0.
65
+
66
+ ---
67
+
68
+ ## Supported tasks
69
+
70
+ - **Host-telemetry anomaly detection.** Per-event or per-window scoring
71
+ against the time-window labels in `labels.csv`.
72
+ - **Sequence anomaly modeling.** Autoregressive next-event prediction (NLL)
73
+ or masked-event modeling (MEM, including field-aware MEM-FA) over the
74
+ integer-coded feature columns.
75
+ - **Event-stream classification baselines.** N-gram, IsolationForest,
76
+ XGBoost, etc. operate directly on the categorical / hash columns.
77
+
78
+ The corpus is **not** intended for: attack-payload synthesis, ground-truth
79
+ event-level labels (only intervals are labeled), or production deployment
80
+ without recalibration to the operator's own host distribution.
81
+
82
+ ---
83
+
84
+ ## Languages
85
+
86
+ N/A — system telemetry, not natural language.
87
+
88
+ ---
89
+
90
+ ## Dataset structure
91
+
92
+ ### Splits
93
+
94
+ | Split | Rows | Time span (UTC)† | Attack events | Attack rate |
95
+ |-------|------------|---------------------------------------------------|---------------|-------------|
96
+ | train | 16,951,950 | 2026-04-14 01:41:13 → 2026-04-19 22:09:18 | 0 | 0.0 % |
97
+ | test | 2,349,965 | 2026-04-19 21:46:27 → 2026-04-20 02:01:27 | 705,852 | 30.0 % |
98
+
99
+ **† Pre-window event timestamps.** ≈295 events (267 train + 28 test, 0.0015 %
100
+ of the corpus) carry `event_time` values that predate the stated collection
101
+ window. All have `f_is_procfs_walk = 1`: parser-synthesized `process_exec`
102
+ records for daemons (`systemd`, `bash`, `dbus-daemon`, `polkitd`, `cron`,
103
+ `sshd`, etc.) that were already running when Tetragon collection started,
104
+ using the daemon's original `process_start_time` as the event timestamp.
105
+ They preserve lineage roots for downstream events. Filter
106
+ `WHERE f_is_procfs_walk = 0` to exclude them from time-window analysis.
107
+
108
+ ### Schema evolution (v0.2.1)
109
+
110
+ The release artifact is the **encoded** parquet (63 columns: 39 model-input
111
+ columns + 7 hash columns + identifier and side-string columns). Compared to
112
+ v0.2.0 the encoded schema gained four 3a′ promoted categoricals
113
+ (`f_path_category`, `f_dst_ip_category`, `f_dst_port_category`,
114
+ `f_object_category`) and lost two constant-zero booleans (`f_in_init_tree`,
115
+ `f_args_truncated`); two hash columns were right-sized
116
+ (`f_proc_name_hash` 1024 → 2048, `f_proc_cwd_hash` 256 → 4096 with depth
117
+ 2 → 4 path components hashed). The full integer-code → label mapping is in
118
+ `reproduction/feature_definitions.json` (schema version 1.1).
119
+
120
+ The release pipeline is three-stage:
121
+
122
+ | Stage | Source field | Stage column | Transformation |
123
+ |-------|-----------------------------------------------|-------------------------------------------|----------------|
124
+ | Raw Tetragon JSON | `process_exec` / `process_exit` / `process_kprobe` envelope | `event_type` | passthrough → vocab `f_event_type` |
125
+ | → Parser (126-col raw, withheld) | `process.binary` | `proc_binary` (full path) | derive basename → `proc_binary_basename` (released, aliased) + hash `f_proc_name_hash` |
126
+ | | `process.cwd` | `proc_cwd` (full path) | normalize → `proc_cwd_normalized` (path-rewritten at release) + hash `f_proc_cwd_hash` |
127
+ | | `process.uid` | `proc_uid` (uint32) | bucket → `f_proc_uid_bucket` (uint8) — raw uid not released |
128
+ | | `process.arguments` | `proc_arguments` | length-bucket → `f_args_length_bucket` (uint8); raw args not released |
129
+ | | `process_kprobe.args[].sock.daddr` | `kp_sock_daddr` | category bucket → `dst_ip_category` (string) — raw IP not released |
130
+ | → Behavior builder (encoded, released) | (event_type, kprobe_function_name) | `f_action_family` (uint8, 15 codes) | derived; see `feature_definitions.json` |
131
+ | | (parent_basename, proc_basename) pair | `f_parent_child_pair_hash` (uint16, 1024 buckets) | hash |
132
+ | | proc_exec_id walk | `f_lineage_depth`, `f_root_ancestor_basename_hash`, `f_process_tree_id_hash` | derived |
133
+
134
+ The **126-column raw parser output is withheld** from the public release.
135
+ All raw paths, raw arguments, raw IPs, raw uids, hostnames, and PID-namespace
136
+ inums are dropped at this stage; see "Personal and sensitive information"
137
+ below for the per-column decisions.
138
+
139
+ ### Feature column inventory (encoded parquet)
140
+
141
+ Full integer-code → label mapping for every categorical column lives in
142
+ `reproduction/feature_definitions.json`. High-level summary:
143
+
144
+ - **22 integer-coded categorical columns** — 18 from v0.2.0 plus the four
145
+ 3a′ promoted categoricals (`f_path_category` 11-way, `f_dst_ip_category`
146
+ 6-way, `f_dst_port_category` 7-way, `f_object_category` 8-way).
147
+ - **4 string-coded categorical columns** (`path_category`, `object_category`,
148
+ `dst_ip_category`, `dst_port_category`) — provenance side columns,
149
+ identical semantics to their `f_*` integer counterparts.
150
+ - **7 hash columns** with declared bucket capacities (2048 / 1024 / 4096 /
151
+ 256 / 1024 / 1024 / 4096 — see *Hash-bucket utilization* below).
152
+ - **6 boolean flag columns** (e.g. `f_is_procfs_walk`, `f_uid_eq_parent`,
153
+ `f_kp_commit_creds_uid_change`).
154
+ - **5 numeric columns** (`delta_t_prev_ns`, `process_age_ns`, mmap/mprotect
155
+ prot bitmasks, fd integer).
156
+ - **3 bucketed log-time columns** (`f_delta_t_log_bucket`,
157
+ `f_process_age_log_bucket`, `f_lineage_depth`).
158
+ - **3 identifier columns** (`event_time`, `proc_exec_id`, `proc_pid`).
159
+ - **7 release-eligible side-string columns** (basenames, normalized cwd,
160
+ parent-child pair, token, process_tree_root_exec_id).
161
+
162
+ ### Class distributions (train split)
163
+
164
+ **`f_event_type`** (4-way; OOV not observed):
165
+
166
+ | Code | Label | Count | % |
167
+ |------|-------------------|-------------|-------|
168
+ | 2 | process_kprobe | 16,413,759 | 96.83 |
169
+ | 1 | process_exit | 288,716 | 1.70 |
170
+ | 0 | process_exec | 249,475 | 1.47 |
171
+
172
+ **`f_action_family`** (15-way; top 10 shown; full table in
173
+ `feature_definitions.json`):
174
+
175
+ | Code | Label | Count |
176
+ |------|--------------------|-------------|
177
+ | 6 | FILE_OPEN | 14,852,286 |
178
+ | 12 | MEM_MAP | 1,168,402 |
179
+ | 1 | PROC_EXIT | 288,716 |
180
+ | 9 | PRIV_CHANGE | 257,709 |
181
+ | 0 | PROC_EXEC | 249,475 |
182
+ | 7 | FILE_DELETE | 44,818 |
183
+ | 5 | NET_SEND | 23,672 |
184
+ | 13 | MEM_PROTECT | 19,059 |
185
+ | 3 | NET_CLOSE | 18,451 |
186
+ | 2 | NET_CONNECT | 16,359 |
187
+
188
+ ### ATT&CK label coverage
189
+
190
+ 61 labeled time intervals span 21 ATT&CK techniques + 1 supply-chain capstone
191
+ chain across 8 tactics. **All attack-positive intervals are in the test
192
+ split.** Per-technique event counts (test split, top 10 by event volume):
193
+
194
+ | ID | Technique | Test events |
195
+ |-------------|------------------------------------------------------------|-------------|
196
+ | T1059.004 | Unix Shell | 175,797 |
197
+ | T1083 | File and Directory Discovery | 154,581 |
198
+ | T1552.001 | Credentials In Files | 97,493 |
199
+ | T1543.002 | Systemd Service | 47,901 |
200
+ | CHAIN.CAP01 | Capstone Chain v0.1 (Supply-Chain Narrative) | 34,130 |
201
+ | T1548.003 | Sudo and Sudo Caching | 24,229 |
202
+ | T1059.006 | Python | 22,675 |
203
+ | T1562.001 | Disable or Modify Tools | 18,610 |
204
+ | T1195.002 | Compromise Software Supply Chain | 15,703 |
205
+ | T1548.001 | Setuid and Setgid | 12,461 |
206
+
207
+ Full per-technique table is reproducible from `notebooks/v0-2_dataset_audit.ipynb`
208
+ (cell 63).
209
+
210
+ ### Hash-bucket utilization (train split, post-3a′ caps)
211
+
212
+ | Column | Capacity | Buckets used | Saturation |
213
+ |---------------------------------|----------|--------------|------------|
214
+ | f_proc_name_hash | 2048 | 269 | 13.1 % |
215
+ | f_parent_proc_hash | 1024 | 121 | 11.8 % |
216
+ | f_proc_cwd_hash | 4096 | 65 | 1.6 % |
217
+ | f_lineage_bag_hash | 256 | 98 | 38.3 % |
218
+ | f_parent_child_pair_hash | 1024 | 490 | 47.9 % |
219
+ | f_root_ancestor_basename_hash | 1024 | 5 | 0.5 % |
220
+ | f_process_tree_id_hash | 4096 | 7 | 0.2 % |
221
+
222
+ Low saturation reflects the single-host single-developer baseline (a small
223
+ real basename / cwd / lineage population). The 3a′ capacity bumps on
224
+ `f_proc_name_hash` and `f_proc_cwd_hash` were sized for plausible multi-host
225
+ saturation rather than this corpus's observed counts; the values are
226
+ documented here so re-trainers can resize the embedding tables.
227
+
228
+ ---
229
+
230
+ ## Dataset creation
231
+
232
+ ### Curation rationale
233
+
234
+ This dataset was assembled to support the v0.2 milestone of an academic
235
+ project investigating Mamba-based anomaly detection on host telemetry. The
236
+ v0.1 predecessor produced an *inverted* AUROC (~0.35) on a similar but
237
+ parser-buggy corpus; v0.2 was rebuilt with a Tetragon-native parser,
238
+ sentinel-window exclusion at parse time, and a two-stage
239
+ parser → behavior-builder design that preserves a raw normalized parquet
240
+ internally for iteration. The **v0.2.1 release** further re-parsed the same
241
+ collection on 2026-04-29 with five correctness fixes:
242
+
243
+ - Procfs-walk synthetic execs are now **tagged, not dropped** (1,346 events
244
+ recovered corpus-wide; lineage gaps for daemons running before agent start
245
+ closed).
246
+ - `delta_t_prev_ns` is keyed **per-process** rather than globally, removing
247
+ 44,947 train + 1,643 test negative residuals.
248
+ - Test-split lineage walking now seeds from the train parquet first,
249
+ resolving 39 % of previously-missing parents (test
250
+ `lineage_missing_parent`: 5.25 % → 2.88 %).
251
+ - Parser parallelized across files via `multiprocessing.Pool`, with DuckDB
252
+ `ORDER BY event_time` for chronological output.
253
+ - Mamba NLL scorer's front-pad alignment fix: `nll_padded[i]` is now event
254
+ *i*'s surprise (was *i+1*'s in v0.2.0).
255
+
256
+ ### Source data
257
+
258
+ - **Sensor:** Tetragon 1.6.1 (Cilium eBPF) running on a Linux host.
259
+ - **Workload:** a single LXC container hosting a developer's daily workspace
260
+ (Node, Python, VSCode, npm). Background services inside the container
261
+ (systemd-resolved, dbus, polkitd, etc.) contribute system noise.
262
+ - **Tracing policies:** 7 YAML policies covering file, network, memory,
263
+ privilege, process-injection, and file-attribute kprobes (bundled in
264
+ `reproduction/tetragon-configs/`).
265
+ - **Attribution:** the parser pins on PID-namespace inum, which uniquely
266
+ identifies the target container across an inum rotation that occurred
267
+ during the collection window. Host-namespace events are dropped at parse
268
+ time.
269
+
270
+ ### Annotations
271
+
272
+ Time-window labels are provided in `labels.csv` (61 rows; 22 distinct
273
+ technique IDs incl. one capstone chain; 8 tactics). Labels were generated
274
+ by an Atomic Red Team campaign run inside the same container during the
275
+ test window, with start/end timestamps captured per-trial. The campaign
276
+ manifest is bundled at `reproduction/art-campaign/`.
277
+
278
+ **Sentinel-window exclusion.** Four short intervals on 2026-04-19
279
+ corresponding to ART setup-shakedown / dry-run / smoke-test bursts (NOT
280
+ labeled-attack telemetry) are excluded by the parser; the audit notebook
281
+ verifies zero events in any of those intervals across both raw and encoded
282
+ parquets.
283
+
284
+ ### Personal and sensitive information
285
+
286
+ The v0.2.1 anonymizer (`scripts/anonymize_v0_2_parquet.py`) is
287
+ **component-first** with deep aliasing across all PII-bearing string
288
+ columns. Summary:
289
+
290
+ - **126-column raw parquet withheld.** All raw paths, raw arguments, raw
291
+ IPs, raw uids, hostnames, and PID-namespace inums are dropped at the
292
+ parser stage. The released parquet is the 63-column encoded form.
293
+ - **Basename aliasing.** A single alias map covers `proc_binary_basename`,
294
+ `parent_binary_basename`, and `root_ancestor_basename`. Distro-standard
295
+ binaries and dev-tooling that anchor the "standardized agentic-research
296
+ workstation" baseline (`bash`, `node`, `python3.12`, `claude`, `gh`,
297
+ `npm`, `code`, `git`, etc.) pass through verbatim. Custom binaries
298
+ (ART trial scripts, build artifacts) become `custom_tool_NNN` (lex-sorted).
299
+ VSCode commit-hash binaries (`code-<40hex>`) collapse to a single
300
+ `vscode_server_bin` alias. The full original → alias map is shipped at
301
+ `reproduction/alias_map.json`.
302
+ - **Composite re-composition.** `parent_child_pair` and `token` are
303
+ regenerated from the aliased atomic columns via the canonical
304
+ `format_token`, so the original (un-aliased) basename never re-enters
305
+ these fields.
306
+ - **Path rewriting.** `proc_cwd_normalized` is rewritten by ordered rules:
307
+ `/home/<user>` → `/home/_user_`, `/Users/<user>` → `/Users/_user_`,
308
+ `/mnt/projects/<X>` → `/workspace/_project_`, `/projects/<X>` →
309
+ `/workspace/_project_`, residual `rypow_` → `custom_`.
310
+ - **Exec-id pseudonymization.** `proc_exec_id` and
311
+ `process_tree_root_exec_id` are rewritten as
312
+ `exec_<32hex>` via salted HMAC-SHA256 (16-byte digest, hex-encoded).
313
+ Equality joins are preserved (lineage walks still work); the original
314
+ base64-encoded `host:ts:pid` payload is erased.
315
+ - **Defensive global string sweep.** A final substring pass over every
316
+ string column substitutes `rypow → research_user`,
317
+ `prox1 → sandbox_host`, `dendroaspis → agent_platform_v02`,
318
+ `ryan_admin → privileged_user` as a safety net.
319
+ - **Post-write assertions.** A regex sweep over the released shards
320
+ (notebook §22-24) confirms zero matches on raw or base64-decoded values
321
+ for any of: `prox1`, `rypow`, `ryan_admin`, `/home/rypow/`,
322
+ `/projects/ml_final_project`, `/mnt/projects/`, `dendroaspis`,
323
+ `/datasets/`, MAC-address pattern. `proc_exec_id` and
324
+ `process_tree_root_exec_id` are format-checked against
325
+ `^exec_[0-9a-f]{32}$`. Every basename in the three basename columns is
326
+ asserted to be in the alias-map closure.
327
+
328
+ ---
329
+
330
+ ## Considerations for using the data
331
+
332
+ ### Sources of bias
333
+
334
+ - **Single host, single developer.** All baseline (benign) telemetry comes
335
+ from one Linux container running one developer's daily workflow over six
336
+ days. Tool habits (VSCode, npm, git, claude-code) appear at high
337
+ frequency; absent activities (CI/CD pipelines, multi-tenant traffic, GPU
338
+ jobs) are simply not represented.
339
+ - **Container-internal UIDs.** Tetragon 1.6.1 emits container-internal UIDs;
340
+ the v0.2 encoder buckets uids into 5 classes (ROOT / RESEARCH_USER /
341
+ SYSTEM_LOW / DAEMON_HIGH / OTHER) rather than treating them as a
342
+ continuous identifier. Researchers reusing these buckets on multi-host
343
+ data should re-bucket.
344
+ - **Time skew.** Train spans 5 d 20 h, test spans only 4 h 15 m. The test
345
+ window concentrates 30 % attack-event mass, which is unrealistically dense
346
+ for production traffic; downstream metrics (AUROC, AP) on this corpus
347
+ reflect the campaign's deliberate density.
348
+ - **Hash-bucket sparsity.** Hash columns are sized for plausible multi-host
349
+ saturation; on this single-host corpus most are < 50 % utilized (see
350
+ hash-bucket table above).
351
+ - **Aliased basename narrative.** The v0.2.1 anonymizer collapses a long
352
+ tail of one-off custom binaries into `custom_tool_NNN` slots and folds
353
+ per-build VSCode commit-hash binaries into `vscode_server_bin`. The
354
+ resulting parquet looks like a "standardized agentic-research workstation"
355
+ baseline more than the specific developer's machine; the true diversity
356
+ of custom-script names was higher than what survives in the released
357
+ string columns.
358
+
359
+ ### Social impact
360
+
361
+ This is a defensive-research dataset. It enables researchers to develop and
362
+ benchmark host-anomaly-detection models without operating live attack
363
+ infrastructure. The bundled ATT&CK technique inventory is **not** an
364
+ attack tutorial — payload generation requires a separate Atomic Red Team
365
+ installation and is out of scope.
366
+
367
+ ### Known limitations
368
+
369
+ - **No ground-truth event-level labels.** Labels are time-window only; some
370
+ benign system activity (e.g. background daemons running during an ART
371
+ technique window) is *labeled-attack* by association.
372
+ - **Train has zero attack events.** All ART intervals fall in test by
373
+ construction. Unsupervised / self-supervised methods are the natural fit;
374
+ fully supervised classifiers would need to engineer their own
375
+ train-time positives.
376
+ - **Single Tetragon version.** Format quirks of Tetragon 1.6.1 (e.g.
377
+ KPROBE_ACTION enum coverage, ns-PID inum semantics) are baked in. Older
378
+ / newer Tetragon emits slightly different envelopes; the parser pin in
379
+ `reproduction/PARSER_VERSION` documents the exact reference build.
380
+ - **Linux only.** Tetragon is Linux-only; this corpus has no Windows or
381
+ macOS analogue.
382
+
383
+ ### Intended use + dual-use disclaimer
384
+
385
+ **Intended use.** This dataset supports:
386
+
387
+ - training and evaluating sequence / anomaly models on host telemetry,
388
+ - benchmarking baselines (n-gram, XGBoost, IsolationForest, LSTM, Mamba,
389
+ Transformer) on a fixed split,
390
+ - studying ATT&CK-technique-specific detectability,
391
+ - research on attack representations, lineage features, and event encoders.
392
+
393
+ **Not intended use.** The attack telemetry is real ATT&CK technique
394
+ execution but generated in an isolated lab LXC container; the corpus is
395
+ **not** an attack template, **not** a payload library, and **not** a
396
+ substitute for production EDR data. Operators who reuse the bundled
397
+ tracing policies + parser to collect their own corpus are responsible for
398
+ their own sandboxing.
399
+
400
+ ---
401
+
402
+ ## Additional information
403
+
404
+ ### License
405
+
406
+ CC BY 4.0 — see `LICENSE` in this repo. Code under
407
+ `reproduction/scripts/` (in the project repository) is MIT-licensed.
408
+
409
+ ### Citation
410
+
411
+ ```bibtex
412
+ @dataset{dendroaspis_v0_2_2026,
413
+ title = {v0.2 Tetragon-native HIDS corpus with 21 ATT\&CK techniques},
414
+ author = {Powers, Ryan William},
415
+ year = {2026},
416
+ version = {0.2.1},
417
+ doi = {<TBD-on-publish>},
418
+ url = {<TBD-on-publish>}
419
+ }
420
+ ```
421
+
422
+ ### Contact
423
+
424
+ GitHub: [@ryypow](https://github.com/ryypow). Issues / corrections
425
+ preferred via the project repository.
datasheet.md ADDED
@@ -0,0 +1,413 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Datasheet — v0.2 Tetragon-native HIDS corpus
2
+
3
+ Following the [Datasheets for Datasets](https://arxiv.org/abs/1803.09010)
4
+ framework (Gebru et al.). Cross-references the
5
+ [`datacard.md`](./datacard.md) for numerical breakdowns; this document
6
+ covers process and provenance.
7
+
8
+ This datasheet describes the **v0.2.1 release** — same Tetragon collection
9
+ as v0.2.0, re-parsed on 2026-04-29 with five correctness fixes plus the
10
+ 3a′ encoder revisions. See `reproduction/PARSER_VERSION` for the full
11
+ changelog.
12
+
13
+ ---
14
+
15
+ ## 1. Motivation
16
+
17
+ ### Why was this dataset created?
18
+
19
+ To support an academic course-milestone project (v0.2) investigating
20
+ distributed Mamba state-space models for host-telemetry anomaly detection.
21
+ The v0.1 predecessor produced an *inverted* test AUROC (~0.35), which
22
+ turned out to be parser-driven; v0.2 was rebuilt with a Tetragon-native
23
+ parser and a two-stage parser → behavior-builder design. A reproducible
24
+ external corpus is the precondition for the v0.2 paper's "we release X"
25
+ claim, for §3 reproducibility, and for the §6 limitations bound to
26
+ "single-host developer-baseline."
27
+
28
+ ### Who created this dataset and on behalf of which entity?
29
+
30
+ Ryan William Powers, in the context of an academic course project. No
31
+ commercial entity is involved; collection happened on the author's own
32
+ Linux host.
33
+
34
+ ### Who funded the creation of the dataset?
35
+
36
+ Self-funded (homelab hardware, no compute / cloud charge).
37
+
38
+ ### Any other comments?
39
+
40
+ The release is contingent on the author's anonymization sweep
41
+ (`scripts/anonymize_v0_2_parquet.py` + the per-column decisions
42
+ documented in the datacard's "Personal and sensitive information" section);
43
+ no third-party identifiable data was captured because no third-party
44
+ traffic crosses the isolated LXC container.
45
+
46
+ ---
47
+
48
+ ## 2. Composition
49
+
50
+ ### What do the instances represent?
51
+
52
+ Each instance is a single **OS event** — one row of the parquet —
53
+ emitted by Tetragon eBPF probes inside an LXC container during normal
54
+ developer activity or during a labeled ATT&CK technique trial. Events
55
+ fall into three top-level types: `process_exec`, `process_exit`, and
56
+ `process_kprobe` (with the kprobe sub-type captured in
57
+ `f_kprobe_function`).
58
+
59
+ ### How many instances are there?
60
+
61
+ - Train: 16,951,950 events (no attack labels by design)
62
+ - Test: 2,349,965 events (705,852 inside labeled-attack windows ≈ 30 %)
63
+ - Total: 19,301,915 events
64
+
65
+ ### Does the dataset contain all possible instances?
66
+
67
+ No. The dataset is a **time-bounded sample** from one host between
68
+ 2026-04-14T01:41Z and 2026-04-20T02:01Z. Only events the parser admits
69
+ (post container filter, post sentinel-window exclusion) reach the
70
+ parquet. Procfs-walk synthetic execs are **tagged** (via
71
+ `f_is_procfs_walk = 1`) rather than dropped, so they are present in the
72
+ release; filter on the flag if needed.
73
+
74
+ ### What does each instance consist of?
75
+
76
+ A 63-column row mixing:
77
+ - 22 integer-coded categoricals (`f_event_type`, `f_action_family`, etc.,
78
+ including the four 3a′ promoted categoricals: `f_path_category`,
79
+ `f_dst_ip_category`, `f_dst_port_category`, `f_object_category`)
80
+ - 4 string-coded categoricals (`path_category`, `object_category`,
81
+ `dst_ip_category`, `dst_port_category` — provenance side columns)
82
+ - 7 hash columns (256–4096 buckets each)
83
+ - 6 boolean flags (`f_in_init_tree` and `f_args_truncated` were dropped
84
+ in 3a′ as constant-zero in this corpus)
85
+ - 5 numeric columns (timestamps in ns, capability bitmasks, fd integer)
86
+ - 3 bucketed log-time columns
87
+ - 3 identifier columns (`event_time`, `proc_exec_id`, `proc_pid`)
88
+ - 7 release-eligible side strings (basenames, normalized cwd,
89
+ `parent_child_pair`, `token`, `process_tree_root_exec_id`)
90
+
91
+ Full schema and integer-code → label mapping in
92
+ `reproduction/feature_definitions.json` (schema version 1.1).
93
+
94
+ ### Is there a label or target associated with each instance?
95
+
96
+ Not directly. **Time-window labels** in `labels.csv` provide
97
+ `(host, technique_id, technique_name, tactic, start_ts, end_ts, trial,
98
+ delivery, notes)` per ART trial. To label an instance, join on
99
+ `event_time >= start_ts AND event_time < end_ts` (half-open interval —
100
+ matches the audit notebook); the event inherits the intersected
101
+ interval's technique label. Events outside any interval are benign by
102
+ exclusion.
103
+
104
+ ### Is any information missing from individual instances?
105
+
106
+ Yes, by design:
107
+
108
+ - Raw paths (`proc_binary`, `proc_cwd`, `parent_*`, `kp_*_path`) are dropped.
109
+ - Raw arguments (`proc_arguments`, `parent_arguments`, `kprobe_args_json`)
110
+ are dropped.
111
+ - Raw IPs (`kp_sock_daddr`, `kp_sock_saddr`) are dropped.
112
+ - Raw uids and namespace inums are dropped.
113
+ - Hostnames (`node_name`) are dropped.
114
+ - `proc_exec_id` and `process_tree_root_exec_id` are pseudonymized to
115
+ `exec_<32hex>` via salted HMAC-SHA256; equality joins are preserved
116
+ (lineage walks still work) but the original `host:ts:pid` payload
117
+ is erased.
118
+ - Basenames (`proc_binary_basename`, `parent_binary_basename`,
119
+ `root_ancestor_basename`) are aliased: distro / dev-tooling pass
120
+ through verbatim, custom binaries become `custom_tool_NNN`, VSCode
121
+ commit-hash binaries collapse to `vscode_server_bin`. Full alias map
122
+ at `reproduction/alias_map.json`.
123
+
124
+ These are reachable in the original raw parquet (not released; see §6
125
+ distribution) and are reconstructible from the bundled parser + tracing
126
+ configs by re-collecting one's own data.
127
+
128
+ ### Are relationships between individual instances made explicit?
129
+
130
+ Yes — process lineage is preserved through three columns:
131
+
132
+ - `proc_exec_id` — pseudonymized but stable per-exec identifier.
133
+ - `f_lineage_depth` + `f_lineage_bag_hash` — encoder-side ancestry.
134
+ - `process_tree_root_exec_id` + `f_root_ancestor_basename_hash` — root
135
+ of the process subtree.
136
+
137
+ Together they let downstream code reconstruct walks up to the process
138
+ tree root. The v0.2.1 rebuild's train→test lineage seeding closed 39 %
139
+ of the previously-missing test parents (5.25 % → 2.88 %).
140
+
141
+ ### Are there recommended data splits?
142
+
143
+ Yes. **Use the chronological train/test split as shipped.** All ART
144
+ attack intervals fall in test; train is benign-only by construction.
145
+ This is the split the v0.2 paper reports on.
146
+
147
+ ### Are there errors / sources of noise?
148
+
149
+ - **Sentinel-window contamination** (small ART setup / dry-run /
150
+ smoke-test bursts) is excluded by the parser and verified by audit
151
+ notebook §15.
152
+ - **`procFS` synthetic execs** (Tetragon walks /proc/ on startup and
153
+ emits one synthetic exec per pre-existing process) are **tagged**
154
+ via `f_is_procfs_walk = 1` rather than dropped; ≈295 of these have
155
+ `event_time` predating the stated collection window because their
156
+ `process_start_time` is the daemon's original start time. Filter
157
+ `f_is_procfs_walk = 0` to exclude.
158
+ - **Container-attribution leakage** is verified absent — raw rows all
159
+ have `proc_ns_pid_inum` in the target-container set.
160
+ - **Clock jumps** (delta_t > 1 h on the same `proc_exec_id`) are
161
+ flagged in the audit notebook and called out in `datacard.md` if any
162
+ survived collection. The v0.2.1 rebuild's per-process `delta_t_prev_ns`
163
+ keying eliminates the prior 44,947 train + 1,643 test negative
164
+ residuals to literal zero.
165
+
166
+ ### Is the dataset self-contained?
167
+
168
+ Almost. The parquet shards + `labels.csv` + `feature_definitions.json`
169
+ are the canonical artifact. `reproduction/` contains the configs and
170
+ parser SHA needed to recollect raw Tetragon data; collection itself
171
+ requires a Linux host running Tetragon 1.6.1 and is not bundled.
172
+
173
+ ### Does the dataset contain confidential / offensive content?
174
+
175
+ No confidential third-party data — collection happened on an isolated
176
+ LXC container with no external user traffic. ATT&CK technique execution
177
+ is by construction adversarial-style activity but happened only against
178
+ the author's own container. The v0.2.1 anonymizer further de-identifies
179
+ the released parquet (basename aliasing, path rewriting, exec-id
180
+ pseudonymization, defensive global string sweep — see
181
+ `datacard.md` "Personal and sensitive information").
182
+
183
+ ---
184
+
185
+ ## 3. Collection process
186
+
187
+ ### How was the data acquired?
188
+
189
+ - **Sensor:** Tetragon 1.6.1 (Cilium eBPF) running as a host-level daemon.
190
+ - **Target:** a single LXC container on the host. The parser pins on the
191
+ container's PID-namespace inum at filter time.
192
+ - **Tracing policies:** 7 YAML policies covering file (fd_install,
193
+ do_unlinkat, chmod_common), network (tcp_connect / close, inet_csk_accept,
194
+ udp_sendmsg), memory (security_mmap_file, security_file_mprotect),
195
+ privilege (commit_creds), process injection (sys_ptrace,
196
+ sys_process_vm_writev), and file-attribute hooks. Bundled in
197
+ `reproduction/tetragon-configs/`.
198
+ - **Output format:** rotated `events-*.json.gz` files written to
199
+ `/var/log/tetragon/` and rsynced to a NAS for archival.
200
+
201
+ ### What mechanisms / procedures collected the data?
202
+
203
+ Tetragon export → rotation → archive → parser. The parser is
204
+ `mamba-edge/src/telemetry/tetragon_native_parser.py` at SHA `03e2204`
205
+ (post-rebuild + 3a′ encoder revisions). The behavior builder
206
+ (`mamba-edge/src/processing/v0_2_behavior_builder.py`) ran on the parser
207
+ output to produce the encoded parquet. The parallel runner
208
+ `scripts/run_v0.2_parser.py` fans the parser across files via
209
+ `multiprocessing.Pool`, with DuckDB `ORDER BY event_time` for
210
+ chronological output.
211
+
212
+ ### Is the data sample of a larger set?
213
+
214
+ Yes — a time-bounded slice of the host's Tetragon stream during the
215
+ collection window. Earlier and later data from the same host exists but
216
+ is not part of this release.
217
+
218
+ ### Who collected the data, and how were they compensated?
219
+
220
+ The author. No compensation; this is an academic course project.
221
+
222
+ ### Over what timeframe was the data collected?
223
+
224
+ 2026-04-14T01:41:13Z (first event in train) → 2026-04-20T02:01:27Z (last
225
+ event in test). 6-day collection window, with a brief Tetragon outage on
226
+ 2026-04-18→19 (documented in the project repo's incident log) that
227
+ triggered a PID-namespace inum rotation captured in the multi-inum
228
+ `target_pid_ns_inum` filter.
229
+
230
+ ### Were any ethical-review processes conducted?
231
+
232
+ No formal IRB. The dataset captures only the author's own activity on
233
+ the author's own hardware.
234
+
235
+ ### Were the individuals notified / did they consent?
236
+
237
+ The dataset captures the author's own session; the author is the sole
238
+ subject and consents to the release (and applied the v0.2.1 anonymization
239
+ sweep documented in the datacard).
240
+
241
+ ---
242
+
243
+ ## 4. Preprocessing / cleaning / labeling
244
+
245
+ ### Was preprocessing / cleaning done?
246
+
247
+ Yes. Two stages:
248
+
249
+ 1. **Parser** (`tetragon_native_parser.py`):
250
+ - Container filter: drops events outside the target PID-namespace
251
+ inum set.
252
+ - Sentinel-window filter: drops events inside four ART setup / dry-run
253
+ / smoke-test intervals on 2026-04-19.
254
+ - Procfs-walk: synthetic execs from Tetragon's startup /proc/ scan
255
+ are **tagged** (`f_is_procfs_walk = 1`), not dropped — they preserve
256
+ lineage roots for daemons that pre-existed the agent.
257
+ - Field normalization: parses RFC3339-ns timestamps to int64 ns,
258
+ hydrates 126 typed columns from the JSONL envelopes.
259
+
260
+ 2. **Behavior builder** (`v0_2_behavior_builder.py`):
261
+ - Lineage walk to populate `f_lineage_*` and `process_tree_*` columns.
262
+ Train→test seeding closes daemon-rooted parents that lived only in
263
+ the train window.
264
+ - Action-family / object-category / path-category / dst-ip-category /
265
+ dst-port-category derivations.
266
+ - Hashing of basenames / lineage bags / parent-child pairs into
267
+ fixed-capacity buckets (3a′ caps: 2048 / 1024 / 4096 / 256 / 1024 /
268
+ 1024 / 4096).
269
+ - Categorical encoding of every kept column to `uint8` / `uint16`.
270
+ - Per-process `delta_t_prev_ns` keying (eliminates global-stream
271
+ ordering artifacts).
272
+
273
+ 3. **Anonymization** (`scripts/anonymize_v0_2_parquet.py`, run after the
274
+ builder): basename aliasing, path rewriting, composite re-composition,
275
+ exec-id salted-HMAC pseudonymization, defensive global string sweep.
276
+ See the datacard "Personal and sensitive information" section.
277
+
278
+ ### Was the "raw" data saved?
279
+
280
+ Yes, internally. The 126-column raw parquet (`raw_train.parquet`,
281
+ `raw_test.parquet`) is not part of the public release because it would
282
+ require column-level scrubbing across 9 path-bearing columns spanning
283
+ 60–95 % of rows. Re-collection from the bundled configs is the supported
284
+ path for raw data.
285
+
286
+ ### Is the preprocessing software available?
287
+
288
+ Yes. Both stages live in the project repository at the parser SHA pinned
289
+ in `reproduction/PARSER_VERSION`. The repo (linked from the
290
+ [Hugging Face dataset card](./datacard.md)) is permissively licensed (MIT).
291
+
292
+ ---
293
+
294
+ ## 5. Uses
295
+
296
+ ### Has the dataset been used for any tasks?
297
+
298
+ Yes:
299
+
300
+ - **Mamba NLL (autoregressive, Run 2)** — next-event surprise scorer with
301
+ the front-pad alignment fix; test AUROC = 0.7411.
302
+ - **Mamba MEM-FA (field-aware masked-event modeling, Run 2)** —
303
+ per-field categorical CE heads on top of the encoded representation;
304
+ test AUROC = 0.7513 (best Mamba result on this corpus).
305
+ - **Classical-baseline bake-off** — n-gram (token-level), IsolationForest,
306
+ XGBoost, all trained on the same encoded feature columns. Best
307
+ classical: IsolationForest (AUROC 0.6658). Mamba's stack adds ≈0.085
308
+ AUROC over the strongest classical. Raw eval JSONs at
309
+ `mamba-edge/artifacts/v0.2/baseline_*/` in the project repo.
310
+
311
+ ### Is there a repository linking to use cases?
312
+
313
+ The v0.2 paper (in `docs/releases/v0.2-course-milestone/report/`) is the
314
+ canonical use case. The audit notebook
315
+ (`mamba-edge/notebooks/v0-2_dataset_audit.ipynb`) demonstrates the
316
+ DuckDB query patterns external researchers can copy.
317
+
318
+ ### What other tasks could this dataset be used for?
319
+
320
+ - ATT&CK-technique-specific detector evaluation
321
+ - Self-supervised pretraining for host-telemetry transformers
322
+ - Lineage / process-tree representation learning
323
+ - Temporal-anomaly-scoring baselines
324
+ - Sequence-tokenization research (the `token` column is a structured
325
+ composite identifier; alternative tokenizations can be derived)
326
+
327
+ ### Is there anything about the dataset's composition that could result in unfair treatment?
328
+
329
+ Single-host, single-developer baseline. Models trained on this corpus
330
+ **will not generalize** to multi-host or multi-tenant environments
331
+ without recalibration. The datacard's "Sources of bias" enumerates the
332
+ specific limitations.
333
+
334
+ ### Are there tasks for which the dataset should NOT be used?
335
+
336
+ - **Production deployment.** Recalibration is mandatory; the
337
+ hash-bucket sizes are right-sized for plausible multi-host saturation
338
+ but observed utilization here is very low — see datacard's
339
+ hash-bucket utilization table.
340
+ - **Generalization claims about HIDS.** A 21-technique single-host
341
+ evaluation is necessary but not sufficient evidence; use it as one
342
+ bench among many, not the bench.
343
+ - **Attack template / red-team training data.** The corpus is event-side
344
+ telemetry, not payloads.
345
+
346
+ ---
347
+
348
+ ## 6. Distribution
349
+
350
+ ### How will the dataset be distributed?
351
+
352
+ Hugging Face Hub — `<user>/dendroaspis-tetragon-hids` (URL filled in at
353
+ publish time). The release bundle is a directory; the parquet shards are
354
+ under `data/`, datacard / datasheet / license / citation at the root,
355
+ and reproduction artifacts under `reproduction/`. SHA256 manifest
356
+ included (`SHA256SUMS`) plus a JSON release manifest (`RELEASE.json`)
357
+ recording the build timestamp + revision SHA + per-file sha256.
358
+
359
+ ### Is there a fee?
360
+
361
+ No. Public, free.
362
+
363
+ ### When will the dataset be distributed?
364
+
365
+ April 2026 (course-milestone window). Initial release coincides with
366
+ the v0.2 paper draft.
367
+
368
+ ### Will the dataset be distributed under a copyright / IP license?
369
+
370
+ CC BY 4.0 (data) + MIT (bundled scripts).
371
+
372
+ ### Have any third parties imposed IP-based or other restrictions?
373
+
374
+ No.
375
+
376
+ ### Do any export controls / regulatory restrictions apply?
377
+
378
+ No. Behavioral telemetry from a homelab; no controlled technologies are
379
+ captured.
380
+
381
+ ---
382
+
383
+ ## 7. Maintenance
384
+
385
+ ### Who is supporting / hosting / maintaining the dataset?
386
+
387
+ Ryan William Powers via the Hugging Face Hub repo. Bug reports, errata,
388
+ and corrections via GitHub issues on the project repository (linked from
389
+ the datacard).
390
+
391
+ ### How can the manager be contacted?
392
+
393
+ GitHub: [@ryypow](https://github.com/ryypow).
394
+
395
+ ### Will the dataset be updated?
396
+
397
+ Possibly. v0.2 is a course-milestone release; if a v0.3 corpus is
398
+ collected (multi-host, longer window, or with added techniques), it will
399
+ be a separate dataset repo so v0.2 remains a stable citation target.
400
+
401
+ ### Will older versions continue to be supported?
402
+
403
+ Yes — v0.1 (the inverted predecessor) is preserved on the project
404
+ repository's `release/v0.1` branch. v0.2.0 (pre-rebuild) is preserved
405
+ at `data/release/v0.2/archive/v0-2_pre-reparse-2026-04-29/`. v0.2.1 is
406
+ the recommended starting point for new work.
407
+
408
+ ### If others want to contribute extensions / build on the dataset?
409
+
410
+ Forking the dataset repo, applying changes, and citing the original is
411
+ fine under CC BY 4.0. For substantive extensions (e.g. a multi-host
412
+ re-collection), open a GitHub issue first to coordinate naming and
413
+ avoid label-schema drift.
dendroaspis.bib ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ @dataset{dendroaspis_v0_2_2026,
2
+ title = {v0.2 Tetragon-native HIDS corpus with 21 ATT\&CK techniques},
3
+ author = {Powers, Ryan William},
4
+ year = {2026},
5
+ month = apr,
6
+ version = {0.2.1},
7
+ doi = {<TBD-on-publish>},
8
+ url = {<TBD-on-publish>},
9
+ note = {Single-host LXC eBPF telemetry, 19.3M events across 22 technique IDs.
10
+ Headline result: Mamba MEM-FA AUROC = 0.751 on held-out test (Run 2,
11
+ post-rebuild + 3a' encoder revisions); see datacard for full numbers.},
12
+ license = {CC BY 4.0}
13
+ }
labels.csv ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ host,technique_id,technique_name,tactic,start_ts,end_ts,trial,delivery,notes
2
+ sandbox_host,CHAIN.CAP01,Capstone Chain v0.1 (Supply-Chain Narrative),execution,2026-04-20T01:59:31.920547970Z,2026-04-20T01:59:45.019566446Z,1,art,
3
+ sandbox_host,T1003.008,/etc/passwd and /etc/shadow,credential_access,2026-04-20T00:27:19.544699604Z,2026-04-20T00:27:21.324864544Z,1,art,
4
+ sandbox_host,T1003.008,/etc/passwd and /etc/shadow,credential_access,2026-04-20T00:27:26.330849154Z,2026-04-20T00:27:28.103071204Z,2,art,
5
+ sandbox_host,T1003.008,/etc/passwd and /etc/shadow,credential_access,2026-04-20T00:27:33.109311865Z,2026-04-20T00:27:34.913462371Z,3,art,
6
+ sandbox_host,T1027,Obfuscated Files or Information,defense_evasion,2026-04-20T00:26:44.362314451Z,2026-04-20T00:26:46.260972910Z,1,art,
7
+ sandbox_host,T1036.005,Match Legitimate Name or Location,defense_evasion,2026-04-20T00:27:01.981472730Z,2026-04-20T00:27:03.743904270Z,1,art,
8
+ sandbox_host,T1053.003,Cron,persistence,2026-04-19T23:21:15.318086413Z,2026-04-19T23:21:17.129137323Z,1,art,
9
+ sandbox_host,T1053.003,Cron,persistence,2026-04-19T23:21:22.134835969Z,2026-04-19T23:21:23.928586108Z,2,art,
10
+ sandbox_host,T1053.003,Cron,persistence,2026-04-19T23:21:28.934815378Z,2026-04-19T23:21:30.715299868Z,3,art,
11
+ sandbox_host,T1057,Process Discovery,discovery,2026-04-20T00:30:24.402125645Z,2026-04-20T00:30:26.221544426Z,1,art,
12
+ sandbox_host,T1059.004,Unix Shell,execution,2026-04-19T23:18:01.681365242Z,2026-04-19T23:18:06.719474616Z,1,art,
13
+ sandbox_host,T1059.004,Unix Shell,execution,2026-04-19T23:18:11.725866322Z,2026-04-19T23:18:20.496957475Z,2,art,
14
+ sandbox_host,T1059.004,Unix Shell,execution,2026-04-19T23:18:25.503038168Z,2026-04-19T23:18:27.525715696Z,3,art,
15
+ sandbox_host,T1059.006,Python,execution,2026-04-19T23:18:43.341638054Z,2026-04-19T23:18:45.368727507Z,1,art,
16
+ sandbox_host,T1059.006,Python,execution,2026-04-19T23:18:50.375362308Z,2026-04-19T23:18:52.376624040Z,2,art,
17
+ sandbox_host,T1059.006,Python,execution,2026-04-19T23:18:57.383084874Z,2026-04-19T23:20:59.334084903Z,3,art,
18
+ sandbox_host,T1070.004,File Deletion,defense_evasion,2026-04-20T00:25:41.526420121Z,2026-04-20T00:25:43.370513123Z,1,art,
19
+ sandbox_host,T1070.004,File Deletion,defense_evasion,2026-04-20T00:25:48.376868194Z,2026-04-20T00:25:50.213997357Z,2,art,
20
+ sandbox_host,T1070.004,File Deletion,defense_evasion,2026-04-20T00:25:55.220340905Z,2026-04-20T00:25:57.134692866Z,3,art,
21
+ sandbox_host,T1082,System Information Discovery,discovery,2026-04-20T00:30:41.953754187Z,2026-04-20T00:30:44.050727668Z,1,art,
22
+ sandbox_host,T1082,System Information Discovery,discovery,2026-04-20T00:30:49.057145569Z,2026-04-20T00:30:51.145539391Z,2,art,
23
+ sandbox_host,T1082,System Information Discovery,discovery,2026-04-20T00:30:56.152341899Z,2026-04-20T00:30:58.262445401Z,3,art,
24
+ sandbox_host,T1083,File and Directory Discovery,discovery,2026-04-20T00:31:13.943771670Z,2026-04-20T00:31:15.858295030Z,1,art,
25
+ sandbox_host,T1083,File and Directory Discovery,discovery,2026-04-20T00:31:20.864889022Z,2026-04-20T00:31:30.438740733Z,2,art,
26
+ sandbox_host,T1083,File and Directory Discovery,discovery,2026-04-20T00:31:35.445243985Z,2026-04-20T00:31:37.269590947Z,3,art,
27
+ sandbox_host,T1098.004,SSH Authorized Keys,persistence,2026-04-20T00:14:12.358631741Z,2026-04-20T00:14:14.020111589Z,1,art,
28
+ sandbox_host,T1105,Ingress Tool Transfer,command_and_control,2026-04-20T00:31:52.969294229Z,2026-04-20T00:31:55.663769554Z,1,art,
29
+ sandbox_host,T1105,Ingress Tool Transfer,command_and_control,2026-04-20T00:32:00.671484771Z,2026-04-20T00:32:03.128654301Z,2,art,
30
+ sandbox_host,T1105,Ingress Tool Transfer,command_and_control,2026-04-20T00:32:08.135220672Z,2026-04-20T00:32:10.559928579Z,3,art,
31
+ sandbox_host,T1195.002,Compromise Software Supply Chain,initial_access,2026-04-19T22:45:43.962209623Z,2026-04-19T22:45:45.235700734Z,1,art,
32
+ sandbox_host,T1195.002,Compromise Software Supply Chain,initial_access,2026-04-19T22:45:51.463755615Z,2026-04-19T22:45:52.756753227Z,2,art,
33
+ sandbox_host,T1195.002,Compromise Software Supply Chain,initial_access,2026-04-19T22:45:58.980795684Z,2026-04-19T22:46:00.263649678Z,3,art,
34
+ sandbox_host,T1195.002,Compromise Software Supply Chain,initial_access,2026-04-19T22:46:06.486450544Z,2026-04-19T22:46:07.774254382Z,4,art,
35
+ sandbox_host,T1195.002,Compromise Software Supply Chain,initial_access,2026-04-19T22:46:13.991062337Z,2026-04-19T22:46:15.276347100Z,5,art,
36
+ sandbox_host,T1195.002,Compromise Software Supply Chain,initial_access,2026-04-19T22:46:21.493778083Z,2026-04-19T22:46:22.784321750Z,6,art,
37
+ sandbox_host,T1195.002,Compromise Software Supply Chain,initial_access,2026-04-19T22:46:29.005948563Z,2026-04-19T22:46:30.278480726Z,7,art,
38
+ sandbox_host,T1195.002,Compromise Software Supply Chain,initial_access,2026-04-19T22:46:36.495816216Z,2026-04-19T22:46:37.785140546Z,8,art,
39
+ sandbox_host,T1222.002,Linux and Mac File and Directory Permissions Modification,defense_evasion,2026-04-20T00:26:12.922543247Z,2026-04-20T00:26:14.846949090Z,1,art,
40
+ sandbox_host,T1222.002,Linux and Mac File and Directory Permissions Modification,defense_evasion,2026-04-20T00:26:19.853198637Z,2026-04-20T00:26:21.783431606Z,2,art,
41
+ sandbox_host,T1222.002,Linux and Mac File and Directory Permissions Modification,defense_evasion,2026-04-20T00:26:26.789544755Z,2026-04-20T00:26:28.721170208Z,3,art,
42
+ sandbox_host,T1543.002,Systemd Service,persistence,2026-04-19T23:50:59.765650410Z,2026-04-19T23:51:26.521462122Z,1,art,
43
+ sandbox_host,T1543.002,Systemd Service,persistence,2026-04-19T23:52:39.894720646Z,2026-04-19T23:54:17.576136538Z,2,art,
44
+ sandbox_host,T1543.002,Systemd Service,persistence,2026-04-19T23:54:22.582266884Z,2026-04-19T23:54:45.571424667Z,3,art,
45
+ sandbox_host,T1546.004,Unix Shell Configuration Modification,persistence,2026-04-19T23:55:01.568631159Z,2026-04-19T23:55:03.406917745Z,1,art,
46
+ sandbox_host,T1546.004,Unix Shell Configuration Modification,persistence,2026-04-19T23:55:08.413312998Z,2026-04-19T23:55:10.202837668Z,2,art,
47
+ sandbox_host,T1546.004,Unix Shell Configuration Modification,persistence,2026-04-19T23:55:15.209673036Z,2026-04-19T23:55:17.010232860Z,3,art,
48
+ sandbox_host,T1548.001,Setuid and Setgid,privilege_escalation,2026-04-20T00:14:29.982832656Z,2026-04-20T00:14:36.862804138Z,1,art,
49
+ sandbox_host,T1548.001,Setuid and Setgid,privilege_escalation,2026-04-20T00:14:41.869002067Z,2026-04-20T00:14:48.769179332Z,2,art,
50
+ sandbox_host,T1548.001,Setuid and Setgid,privilege_escalation,2026-04-20T00:14:53.775331564Z,2026-04-20T00:14:55.585905430Z,3,art,
51
+ sandbox_host,T1548.003,Sudo and Sudo Caching,privilege_escalation,2026-04-20T00:15:11.314148281Z,2026-04-20T00:17:13.146519169Z,1,art,
52
+ sandbox_host,T1548.003,Sudo and Sudo Caching,privilege_escalation,2026-04-20T00:17:18.152866405Z,2026-04-20T00:17:19.949403438Z,2,art,
53
+ sandbox_host,T1548.003,Sudo and Sudo Caching,privilege_escalation,2026-04-20T00:17:24.955045263Z,2026-04-20T00:17:30.215813570Z,3,art,
54
+ sandbox_host,T1552.001,Credentials In Files,credential_access,2026-04-20T00:27:50.629812895Z,2026-04-20T00:27:52.498058179Z,1,art,
55
+ sandbox_host,T1552.001,Credentials In Files,credential_access,2026-04-20T00:27:57.504584495Z,2026-04-20T00:29:59.576360961Z,2,art,
56
+ sandbox_host,T1552.001,Credentials In Files,credential_access,2026-04-20T00:30:04.583017711Z,2026-04-20T00:30:08.574646575Z,3,art,
57
+ sandbox_host,T1562.001,Disable or Modify Tools,defense_evasion,2026-04-20T00:32:26.379801715Z,2026-04-20T00:33:17.755734382Z,1,art,
58
+ sandbox_host,T1562.001,Disable or Modify Tools,defense_evasion,2026-04-20T00:33:22.761738237Z,2026-04-20T00:33:25.101641640Z,2,art,
59
+ sandbox_host,T1562.001,Disable or Modify Tools,defense_evasion,2026-04-20T00:33:30.108255228Z,2026-04-20T00:33:32.422829942Z,3,art,
60
+ sandbox_host,T1574.006,Dynamic Linker Hijacking,privilege_escalation,2026-04-20T00:19:24.001827850Z,2026-04-20T00:19:25.776246404Z,1,art,
61
+ sandbox_host,T1574.006,Dynamic Linker Hijacking,privilege_escalation,2026-04-20T00:19:30.782463207Z,2026-04-20T00:19:32.576378213Z,2,art,
62
+ sandbox_host,CHAIN.CAP01,Capstone Chain v0.1 (Supply-Chain Narrative),execution,2026-04-20T00:49:29.090978396Z,2026-04-20T00:49:37.045509190Z,2,art,broken 4-atomic chain: T1057:1 -> T1003.008:1 -> T1105:27 -> T1070.004:1; T1059.004 skipped due to invocation arg error (trial arg missing)
reproduction/PARSER_VERSION ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ v0.2 parser pin (post-rebuild, post-3a')
2
+ parser_sha: 03e2204 # 3a' encoder revisions + cwd-hash depth fix + CLI defaults
3
+ release_parser_sha: 03e2204 # commit that materialized the released parquets (2026-04-29 rebuild)
4
+ parser_path: mamba-edge/src/telemetry/tetragon_native_parser.py
5
+ behavior_builder_path: mamba-edge/src/processing/v0_2_behavior_builder.py
6
+
7
+ Changelog vs prior pin (52badf6 / 163c51e):
8
+ - 208bfd3 feat(v0.2): rebuild corpus pipeline (delta_t, sort, alignment, lineage)
9
+ * Parser: procfs-walk events tagged not dropped (1,346 events recovered).
10
+ * Builder: per-process delta_t_prev_ns keying (eliminates 44,947 train + 1,643 test negatives → 0).
11
+ * Builder: train→test lineage seeding (closes 39% of test missing-parent gap).
12
+ * Tooling: parallel parser + DuckDB ORDER BY for chronological output.
13
+ * Scorer: NLL front-pad alignment fix (nll_padded[i] is now event i's surprise).
14
+ - 03e2204 feat(v0.2): apply 3a' encoder revisions + cwd-hash depth fix + CLI defaults
15
+ * Encoder: hash right-sizing on f_proc_name_hash / f_parent_proc_hash / f_proc_cwd_hash.
16
+ * Encoder: cwd depth 2→4 with LIBRARY/USR_BIN/OPT path sub-buckets.
17
+ * Encoder: 4 categoricals promoted to model input (f_path_category, f_dst_ip_category,
18
+ f_dst_port_category, f_object_category).
19
+ * Encoder: f_in_init_tree dropped; f_is_procfs_walk retained.
reproduction/REPRODUCE.md ADDED
@@ -0,0 +1,216 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Reproducing the v0.2 Tetragon HIDS corpus
2
+
3
+ This package contains everything needed to reproduce the *collection +
4
+ parser + encoder* pipeline that produced the released parquet shards. It
5
+ does **not** include the Mamba training pipeline (separate `requirements`
6
+ + GPU); see the project repository's `mamba-edge/src/training/` if you
7
+ want to retrain.
8
+
9
+ ## What's in this directory
10
+
11
+ ```
12
+ reproduction/
13
+ ├── REPRODUCE.md # this file
14
+ ├── PARSER_VERSION # git SHA pin for the parser + builder used
15
+ ├── requirements.txt # minimal pinned freeze for collection / parsing
16
+ ├── feature_definitions.json # integer-code → label tables for every categorical/bucket column
17
+ ├── alias_map.json # original → aliased basename map applied by the anonymizer
18
+ ├── allowlist_basenames.txt # the realised PASS set (basenames preserved verbatim)
19
+ ├── tetragon-configs/
20
+ │ ├── tetragon_parser.yaml # parser runtime config (host / inum / salt slots scrubbed)
21
+ │ ├── identity_map.yaml # uid → persona mapping (placeholders for primary user)
22
+ │ └── tetragon/
23
+ │ ├── file-monitor.yaml # 6 tracing policies covering file / net / mem / priv / inject hooks
24
+ │ ├── network-monitor.yaml
25
+ │ ├── memory-monitor.yaml
26
+ │ ├── privilege-monitor.yaml
27
+ │ ├── injection-monitor.yaml
28
+ │ ├── attribute-monitor.yaml
29
+ │ └── logrotate.tetragon # logrotate config for /var/log/tetragon/
30
+ └── art-campaign/
31
+ ├── ARTcampaign.md # campaign grounding doc (techniques, trial structure)
32
+ └── labels.csv # per-trial (start_ts, end_ts, technique_id, ...) labels
33
+ ```
34
+
35
+ ## Prerequisites
36
+
37
+ - **Linux host** running a kernel that supports Tetragon eBPF (≥ 5.4).
38
+ - **Tetragon 1.6.1** installed and running. Newer Tetragon versions emit
39
+ slightly different envelopes; the parser pinned in `PARSER_VERSION` was
40
+ validated against 1.6.1 specifically.
41
+ - **Container.** The v0.2 corpus targets a single LXC container (CT 110)
42
+ on Proxmox. Any container runtime works; what matters is that the
43
+ parser's PID-namespace-inum filter sees the right inum.
44
+ - **Python 3.12** (for the parser + behavior builder).
45
+
46
+ ## Step-by-step
47
+
48
+ ### Step 1 — Configure Tetragon
49
+
50
+ Copy the bundled tracing policies to your Tetragon installation:
51
+
52
+ ```bash
53
+ sudo mkdir -p /etc/tetragon/tetragon.tp.d
54
+ sudo cp tetragon-configs/tetragon/*.yaml /etc/tetragon/tetragon.tp.d/
55
+ sudo systemctl restart tetragon
56
+ ```
57
+
58
+ Verify Tetragon is emitting events:
59
+
60
+ ```bash
61
+ sudo tail -f /var/log/tetragon/tetragon.log | head -5
62
+ ```
63
+
64
+ ### Step 2 — Configure the parser
65
+
66
+ Edit `tetragon-configs/tetragon_parser.yaml`. The bundled file has
67
+ placeholders for host- and container-specific values:
68
+
69
+ | Field | Replace with |
70
+ |--------------------------|---------------------------------------------|
71
+ | `host_id` | a string identifying your host |
72
+ | `target_pid_ns_inum` | your container's PID-namespace inum (see below) |
73
+ | `user_id_salt` | a per-deployment secret (env var preferred) |
74
+
75
+ To capture your container's PID-ns inum:
76
+
77
+ ```bash
78
+ sudo zcat $(ls -t /var/log/tetragon/events-*.json.gz | head -1) \
79
+ | python3 -c "import json, sys
80
+ for l in sys.stdin:
81
+ d = json.loads(l)
82
+ if 'process_exec' in d:
83
+ print(d['process_exec']['process']['ns']['pid'].get('inum'))
84
+ break"
85
+ ```
86
+
87
+ Re-capture if the container is restarted; the inum changes per boot.
88
+
89
+ ### Step 3 — Collect raw events
90
+
91
+ Tetragon writes rotated `events-*.json.gz` files to `/var/log/tetragon/`.
92
+ Decide a collection window (the v0.2 corpus uses 6 days). Optionally rsync
93
+ the archive to a NAS.
94
+
95
+ ### Step 4 — Run the parser → 126-column raw parquet
96
+
97
+ The parser source lives in the project repository at
98
+ `mamba-edge/src/telemetry/tetragon_native_parser.py` (SHA `03e2204` per
99
+ `PARSER_VERSION`). The recommended runner is the parallel orchestrator
100
+ at `mamba-edge/scripts/run_v0.2_parser.py`, which fans out per-file
101
+ shards across `multiprocessing.Pool` workers and uses DuckDB
102
+ `ORDER BY event_time` for chronological output.
103
+
104
+ From the project repo's root:
105
+
106
+ ```bash
107
+ pip install -r requirements.txt
108
+ V02_PARSER_WORKERS=8 python mamba-edge/scripts/run_v0.2_parser.py \
109
+ --config /path/to/tetragon_parser.yaml \
110
+ --input /var/log/tetragon/ \
111
+ --output data/processed/v0.2/raw_train.parquet
112
+ ```
113
+
114
+ Knobs (env vars):
115
+ - `V02_PARSER_WORKERS=N` — pool size (default 8).
116
+ - `V02_PARSER_RESUME=1` — skip shards whose `.stats.json` sidecar already
117
+ exists (resumable parse).
118
+ - `V02_DUCKDB_MEMORY_LIMIT=8GB` — DuckDB memory cap (auto-detect lies
119
+ inside container memory cgroups; pin this explicitly).
120
+
121
+ Repeat for the test split (different `--input` directory / time-range
122
+ filter).
123
+
124
+ ### Step 5 — Run the behavior builder → 59-column encoded parquet
125
+
126
+ ```bash
127
+ python -m src.processing.v0_2_behavior_builder \
128
+ --in data/processed/v0.2/raw_train.parquet \
129
+ --out data/processed/v0.2/train.parquet
130
+ ```
131
+
132
+ Repeat for the test split.
133
+
134
+ ### Step 6 — (Optional) Apply the same anonymization as the released corpus
135
+
136
+ ```bash
137
+ python mamba-edge/scripts/anonymize_v0_2_parquet.py
138
+ ```
139
+
140
+ The anonymizer is **component-first**: it discovers all unique basenames
141
+ across both splits, builds a single alias map (allowlist of standard
142
+ distro + dev-tool basenames passes through verbatim; anything else
143
+ becomes `custom_tool_NNN` by lex sort, with `code-<40hex>` collapsing to
144
+ `vscode_server_bin`), then applies the map to the three basename
145
+ columns, regenerates `parent_child_pair` and `token` from the aliased
146
+ atomics via the canonical `format_token`, salted-HMAC-hashes
147
+ `proc_exec_id` and `process_tree_root_exec_id` to `exec_<32hex>`,
148
+ rewrites `proc_cwd_normalized` via path rules, and finally applies a
149
+ defensive global string sweep (`rypow → research_user`,
150
+ `prox1 → sandbox_host`, `dendroaspis → agent_platform_v02`,
151
+ `ryan_admin → privileged_user`) across every string column.
152
+
153
+ Post-write, the script verifies zero PII hits across the §22 pattern
154
+ set on every string column — including base64-decoded values, which
155
+ catches host fragments embedded in the original exec_id encoding.
156
+
157
+ Outputs:
158
+ - `data/release/v0.2/data/{train,test}-NNNNN-of-MMMMM.parquet` — sharded.
159
+ - `data/release/v0.2/.salt` — 32-byte HMAC salt; persisted locally for
160
+ re-run reproducibility, **not** included in the public bundle.
161
+ - `data/release/v0.2/reproduction/alias_map.json` — full alias map.
162
+ - `data/release/v0.2/reproduction/allowlist_basenames.txt` — realised
163
+ PASS set.
164
+
165
+ ### Step 7 — (Optional) Reproduce the ATT&CK label timeline
166
+
167
+ `art-campaign/ARTcampaign.md` documents the technique inventory; rerunning
168
+ the campaign requires:
169
+
170
+ - An installation of [Atomic Red Team](https://github.com/redcanaryco/atomic-red-team)
171
+ inside the same container.
172
+ - A driver script that runs each technique with paired
173
+ `mkdir /tmp/art_start_<technique>_<trial>` /
174
+ `mkdir /tmp/art_end_<technique>_<trial>` markers; the parser uses these
175
+ to build labels at parse time.
176
+ - Per-trial timestamp logging into a labels.csv-shaped output.
177
+
178
+ The exact driver script lives in the project repository (out of scope for
179
+ this reproduction package); the `labels.csv` shipped in this directory is
180
+ the canonical timeline and can be used directly without re-running ART.
181
+
182
+ ## Quick sanity check (no recollection needed)
183
+
184
+ If you just want to load the released parquet and verify the schema /
185
+ label join:
186
+
187
+ ```python
188
+ from datasets import load_dataset
189
+ import pandas as pd
190
+
191
+ ds = load_dataset("<user>/dendroaspis-tetragon-hids", streaming=True)
192
+ first = next(iter(ds["train"]))
193
+ print({k: type(v).__name__ for k, v in first.items()})
194
+
195
+ labels = pd.read_csv("labels.csv", parse_dates=["start_ts", "end_ts"])
196
+ labels["start_ts"] = labels["start_ts"].dt.tz_localize("UTC")
197
+ labels["end_ts"] = labels["end_ts"].dt.tz_localize("UTC")
198
+ print(f"{len(labels)} labeled intervals across {labels['technique_id'].nunique()} techniques")
199
+ ```
200
+
201
+ ## What's *not* in this package
202
+
203
+ - The Mamba training pipeline (`mamba-edge/src/training/`) and
204
+ baselines — those need GPU + extra deps (`torch`, `mamba-ssm`,
205
+ `xgboost`, etc.).
206
+ - The raw 126-column parquet — see the dataset card "anonymization
207
+ decisions" section. Re-collection is the supported path.
208
+ - Pretrained model weights and trained-run evaluation artifacts — see
209
+ the dataset card for the canonical headline numbers and the project
210
+ repository's `mamba-edge/artifacts/v0.2/` for raw eval JSONs.
211
+
212
+ ## Versioning
213
+
214
+ This release is **v0.2**. Future releases (v0.3+) will live in separate
215
+ HF dataset repos to keep v0.2 stable as a citation target. See the
216
+ dataset card for a citation block.
reproduction/alias_map.json ADDED
@@ -0,0 +1,333 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "00-header": "00-header",
3
+ "10-help-text": "10-help-text",
4
+ "3": "custom_tool_001",
5
+ "30-systemd-environment-d-generator": "30-systemd-environment-d-generator",
6
+ "50-motd-news": "50-motd-news",
7
+ "9": "custom_tool_002",
8
+ "91-contract-ua-esm-status": "91-contract-ua-esm-status",
9
+ "91-release-upgrade": "91-release-upgrade",
10
+ "92-unattended-upgrades": "92-unattended-upgrades",
11
+ "Alice": "custom_tool_003",
12
+ "Bob": "custom_tool_004",
13
+ "CA": "custom_tool_005",
14
+ "Join": "custom_tool_006",
15
+ "Leave": "custom_tool_007",
16
+ "Merge": "custom_tool_008",
17
+ "Refresh": "custom_tool_009",
18
+ "Setup": "custom_tool_010",
19
+ "VerifyingYourHORSSolution.sh": "custom_tool_011",
20
+ "VerifyingYourSolutionTGDH.sh": "custom_tool_012",
21
+ "agetty": "agetty",
22
+ "apport": "apport",
23
+ "appstreamcli": "appstreamcli",
24
+ "apt": "apt",
25
+ "apt-compat": "apt-compat",
26
+ "apt-config": "apt-config",
27
+ "apt-extracttemplates": "apt-extracttemplates",
28
+ "apt-get": "apt-get",
29
+ "apt-helper": "apt-helper",
30
+ "apt-key": "apt-key",
31
+ "apt.systemd.daily": "apt.systemd.daily",
32
+ "arch": "arch",
33
+ "as": "as",
34
+ "assert_clean_state.sh": "custom_tool_013",
35
+ "awk": "awk",
36
+ "base64": "base64",
37
+ "basename": "basename",
38
+ "bash": "bash",
39
+ "bibtex": "bibtex",
40
+ "browser.sh": "custom_tool_014",
41
+ "c++": "c++",
42
+ "cap": "cap",
43
+ "cat": "cat",
44
+ "cc": "cc",
45
+ "cc1": "cc1",
46
+ "cc1plus": "cc1plus",
47
+ "chattr": "chattr",
48
+ "check-new-release": "check-new-release",
49
+ "chmod": "chmod",
50
+ "chown": "chown",
51
+ "chrome-headless-shell": "chrome-headless-shell",
52
+ "claude": "claude",
53
+ "claude.exe": "claude.exe",
54
+ "cleanup": "cleanup",
55
+ "clear": "clear",
56
+ "clear_console": "clear_console",
57
+ "cloudcode_cli": "cloudcode_cli",
58
+ "cmp": "cmp",
59
+ "cnf-update-db": "cnf-update-db",
60
+ "code": "code",
61
+ "code-41dd792b5e652393e7787322889ed5fdc58bd75b": "vscode_server_bin",
62
+ "code-560a9dba96f961efea7b1612916f89e5d5d4d679": "vscode_server_bin",
63
+ "code-server": "code-server",
64
+ "collect2": "collect2",
65
+ "command-not-found": "command-not-found",
66
+ "containerd": "containerd",
67
+ "cp": "cp",
68
+ "cpuUsage.sh": "custom_tool_015",
69
+ "cron": "cron",
70
+ "crontab": "crontab",
71
+ "curl": "curl",
72
+ "cut": "cut",
73
+ "dash": "dash",
74
+ "date": "date",
75
+ "dbus-daemon": "dbus-daemon",
76
+ "dbus-send": "dbus-send",
77
+ "dbus.postinst": "dbus.postinst",
78
+ "deb-systemd-invoke": "deb-systemd-invoke",
79
+ "debian-sa1": "debian-sa1",
80
+ "diff": "diff",
81
+ "dircolors": "dircolors",
82
+ "dirname": "dirname",
83
+ "dmidecode": "dmidecode",
84
+ "docker": "docker",
85
+ "dpkg": "dpkg",
86
+ "dpkg-db-backup": "dpkg-db-backup",
87
+ "dpkg-deb": "dpkg-deb",
88
+ "dpkg-maintscript-helper": "dpkg-maintscript-helper",
89
+ "dpkg-preconfigure": "dpkg-preconfigure",
90
+ "dpkg-query": "dpkg-query",
91
+ "dpkg-split": "dpkg-split",
92
+ "dpkg-statoverride": "dpkg-statoverride",
93
+ "du": "du",
94
+ "echo": "echo",
95
+ "ed": "ed",
96
+ "env": "env",
97
+ "envsubst": "envsubst",
98
+ "exe": "custom_tool_016",
99
+ "expr": "expr",
100
+ "file": "file",
101
+ "find": "find",
102
+ "findmnt": "findmnt",
103
+ "flock": "flock",
104
+ "friendly-recovery": "friendly-recovery",
105
+ "g++": "g++",
106
+ "gcc": "gcc",
107
+ "gdbus": "gdbus",
108
+ "gemini": "gemini",
109
+ "get_PreReqs.sh": "custom_tool_017",
110
+ "getconf": "getconf",
111
+ "getent": "getent",
112
+ "getopt": "getopt",
113
+ "gettext": "gettext",
114
+ "gh": "gh",
115
+ "git": "git",
116
+ "git-remote-https": "git-remote-https",
117
+ "git-submodule": "git-submodule",
118
+ "gpg-connect-agent": "gpg-connect-agent",
119
+ "gpgconf": "gpgconf",
120
+ "gpgv": "gpgv",
121
+ "grep": "grep",
122
+ "groups": "groups",
123
+ "gtk-update-icon-cache": "gtk-update-icon-cache",
124
+ "gzip": "gzip",
125
+ "head": "head",
126
+ "helpztags": "helpztags",
127
+ "hicolor-icon-theme.postinst": "hicolor-icon-theme.postinst",
128
+ "hostname": "hostname",
129
+ "htop": "htop",
130
+ "http": "http",
131
+ "https": "https",
132
+ "id": "id",
133
+ "install": "install",
134
+ "ip": "ip",
135
+ "jq": "jq",
136
+ "kmod": "kmod",
137
+ "ld": "ld",
138
+ "ld-linux-x86-64.so.2": "ld-linux-x86-64.so.2",
139
+ "ldconfig": "ldconfig",
140
+ "ldconfig.real": "ldconfig.real",
141
+ "ldd": "ldd",
142
+ "less": "less",
143
+ "lesspipe": "lesspipe",
144
+ "libc-bin.postinst": "libc-bin.postinst",
145
+ "ln": "ln",
146
+ "local": "local",
147
+ "locale": "locale",
148
+ "locale-check": "locale-check",
149
+ "login": "login",
150
+ "logrotate": "logrotate",
151
+ "ls": "ls",
152
+ "lsb_release": "lsb_release",
153
+ "lsmod": "lsmod",
154
+ "lspci": "lspci",
155
+ "lxc-start": "lxc-start",
156
+ "lxc-start (deleted)": "lxc-start (deleted)",
157
+ "make": "make",
158
+ "man-db": "man-db",
159
+ "man-db.postinst": "man-db.postinst",
160
+ "mandb": "mandb",
161
+ "master": "master",
162
+ "memfd:lxc-attach (deleted)": "memfd:lxc-attach (deleted)",
163
+ "mkdir": "mkdir",
164
+ "mktemp": "mktemp",
165
+ "mmdc": "mmdc",
166
+ "mount": "mount",
167
+ "mv": "mv",
168
+ "nano": "nano",
169
+ "netplan": "netplan",
170
+ "node": "node",
171
+ "node-gyp": "node-gyp",
172
+ "npm": "npm",
173
+ "nvidia-modprobe": "nvidia-modprobe",
174
+ "nvidia-smi": "nvidia-smi",
175
+ "packagekitd": "packagekitd",
176
+ "pager": "pager",
177
+ "pdfinfo": "pdfinfo",
178
+ "pdflatex": "pdflatex",
179
+ "pet": "pet",
180
+ "pgrep": "pgrep",
181
+ "pickup": "pickup",
182
+ "ping": "ping",
183
+ "pip": "pip",
184
+ "pkill": "pkill",
185
+ "pkttyagent": "pkttyagent",
186
+ "polkit-agent-helper-1": "polkit-agent-helper-1",
187
+ "polkitd": "polkitd",
188
+ "polkitd.postinst": "polkitd.postinst",
189
+ "polkitd.postrm": "polkitd.postrm",
190
+ "polkitd.prerm": "polkitd.prerm",
191
+ "postconf": "postconf",
192
+ "postdrop": "postdrop",
193
+ "postfix-instance-generator": "postfix-instance-generator",
194
+ "prebuild-install": "prebuild-install",
195
+ "preinst": "preinst",
196
+ "printenv": "printenv",
197
+ "printf": "printf",
198
+ "ps": "ps",
199
+ "pwsh": "pwsh",
200
+ "pyenv": "pyenv",
201
+ "pyenv-commands": "pyenv-commands",
202
+ "pyenv-exec": "pyenv-exec",
203
+ "pyenv-hooks": "pyenv-hooks",
204
+ "pyenv-init": "pyenv-init",
205
+ "pyenv-latest": "pyenv-latest",
206
+ "pyenv-prefix": "pyenv-prefix",
207
+ "pyenv-rehash": "pyenv-rehash",
208
+ "pyenv-sh-activate": "pyenv-sh-activate",
209
+ "pyenv-version-file": "pyenv-version-file",
210
+ "pyenv-version-file-read": "pyenv-version-file-read",
211
+ "pyenv-version-name": "pyenv-version-name",
212
+ "pyenv-versions": "pyenv-versions",
213
+ "pyenv-virtualenv-init": "pyenv-virtualenv-init",
214
+ "pyenv-virtualenv-prefix": "pyenv-virtualenv-prefix",
215
+ "pyenv-which": "pyenv-which",
216
+ "python": "python",
217
+ "python3": "python3",
218
+ "python3.12": "python3.12",
219
+ "qmgr": "qmgr",
220
+ "readlink": "readlink",
221
+ "realpath": "realpath",
222
+ "release-upgrade-motd": "release-upgrade-motd",
223
+ "reset": "reset",
224
+ "rg": "rg",
225
+ "rm": "rm",
226
+ "rmdir": "rmdir",
227
+ "rsyslogd": "rsyslogd",
228
+ "ruff": "ruff",
229
+ "run-parts": "run-parts",
230
+ "run-trial.sh": "custom_tool_018",
231
+ "run_atomic_chain.sh": "custom_tool_019",
232
+ "run_atomic_trial.sh": "custom_tool_020",
233
+ "sa1": "sa1",
234
+ "sa2": "sa2",
235
+ "sadc": "sadc",
236
+ "sar.sysstat": "sar.sysstat",
237
+ "savelog": "savelog",
238
+ "scp": "scp",
239
+ "sed": "sed",
240
+ "sendmail": "sendmail",
241
+ "service": "service",
242
+ "setcap": "setcap",
243
+ "setpriv": "setpriv",
244
+ "setup-fake-secrets.sh": "custom_tool_021",
245
+ "sgml-base.postinst": "sgml-base.postinst",
246
+ "sh": "sh",
247
+ "sha256sum": "sha256sum",
248
+ "shred": "shred",
249
+ "sleep": "sleep",
250
+ "snap": "snap",
251
+ "snapd": "snapd",
252
+ "snapd-env-generator": "snapd-env-generator",
253
+ "snapd-generator": "snapd-generator",
254
+ "sort": "sort",
255
+ "ssh": "ssh",
256
+ "sshd": "sshd",
257
+ "sshd-socket-generator": "sshd-socket-generator",
258
+ "stat": "stat",
259
+ "store": "store",
260
+ "stty": "stty",
261
+ "sudo": "sudo",
262
+ "sysstat": "sysstat",
263
+ "systemctl": "systemctl",
264
+ "systemd": "systemd",
265
+ "systemd-cryptsetup-generator": "systemd-cryptsetup-generator",
266
+ "systemd-debug-generator": "systemd-debug-generator",
267
+ "systemd-executor": "systemd-executor",
268
+ "systemd-executor (deleted)": "systemd-executor (deleted)",
269
+ "systemd-fstab-generator": "systemd-fstab-generator",
270
+ "systemd-getty-generator": "systemd-getty-generator",
271
+ "systemd-hibernate-resume-generator": "systemd-hibernate-resume-generator",
272
+ "systemd-integritysetup-generator": "systemd-integritysetup-generator",
273
+ "systemd-journald": "systemd-journald",
274
+ "systemd-logind": "systemd-logind",
275
+ "systemd-logind (deleted)": "systemd-logind (deleted)",
276
+ "systemd-networkd": "systemd-networkd",
277
+ "systemd-networkd-wait-online": "systemd-networkd-wait-online",
278
+ "systemd-rc-local-generator": "systemd-rc-local-generator",
279
+ "systemd-resolved": "systemd-resolved",
280
+ "systemd-run-generator": "systemd-run-generator",
281
+ "systemd-system-update-generator": "systemd-system-update-generator",
282
+ "systemd-sysusers": "systemd-sysusers",
283
+ "systemd-sysv-generator": "systemd-sysv-generator",
284
+ "systemd-timedated": "systemd-timedated",
285
+ "systemd-tmpfiles": "systemd-tmpfiles",
286
+ "systemd-tty-ask-password-agent": "systemd-tty-ask-password-agent",
287
+ "systemd-user-runtime-dir": "systemd-user-runtime-dir",
288
+ "systemd-veritysetup-generator": "systemd-veritysetup-generator",
289
+ "systemd-xdg-autostart-generator": "systemd-xdg-autostart-generator",
290
+ "tail": "tail",
291
+ "tar": "tar",
292
+ "tee": "tee",
293
+ "test": "test",
294
+ "timeout": "timeout",
295
+ "tmux": "tmux",
296
+ "touch": "touch",
297
+ "tr": "tr",
298
+ "trivial-rewrite": "trivial-rewrite",
299
+ "tty": "tty",
300
+ "ubuntu-distro-info": "ubuntu-distro-info",
301
+ "uname": "uname",
302
+ "unattended-upgrade": "unattended-upgrade",
303
+ "uniq": "uniq",
304
+ "unzip": "unzip",
305
+ "update-alternatives": "update-alternatives",
306
+ "update-catalog": "update-catalog",
307
+ "update-icon-caches": "update-icon-caches",
308
+ "update-motd-unattended-upgrades": "update-motd-unattended-upgrades",
309
+ "update-xmlcatalog": "update-xmlcatalog",
310
+ "uptime": "uptime",
311
+ "vim": "vim",
312
+ "vim-common.postinst": "vim-common.postinst",
313
+ "vim-common.postrm": "vim-common.postrm",
314
+ "vim-runtime.postinst": "vim-runtime.postinst",
315
+ "vim-runtime.postrm": "vim-runtime.postrm",
316
+ "vim-tiny.postinst": "vim-tiny.postinst",
317
+ "vim-tiny.prerm": "vim-tiny.prerm",
318
+ "vim.postinst": "vim.postinst",
319
+ "vim.prerm": "vim.prerm",
320
+ "visudo": "visudo",
321
+ "vsce-sign": "vsce-sign",
322
+ "wc": "wc",
323
+ "wget": "wget",
324
+ "which": "which",
325
+ "whoami": "whoami",
326
+ "whois": "whois",
327
+ "xargs": "xargs",
328
+ "xdpyinfo": "xdpyinfo",
329
+ "xrandr": "xrandr",
330
+ "xxd": "xxd",
331
+ "zcat": "zcat",
332
+ "zgrep": "zgrep"
333
+ }
reproduction/allowlist_basenames.txt ADDED
@@ -0,0 +1,308 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 00-header
2
+ 10-help-text
3
+ 30-systemd-environment-d-generator
4
+ 50-motd-news
5
+ 91-contract-ua-esm-status
6
+ 91-release-upgrade
7
+ 92-unattended-upgrades
8
+ agetty
9
+ apport
10
+ appstreamcli
11
+ apt
12
+ apt-compat
13
+ apt-config
14
+ apt-extracttemplates
15
+ apt-get
16
+ apt-helper
17
+ apt-key
18
+ apt.systemd.daily
19
+ arch
20
+ as
21
+ awk
22
+ base64
23
+ basename
24
+ bash
25
+ bibtex
26
+ c++
27
+ cap
28
+ cat
29
+ cc
30
+ cc1
31
+ cc1plus
32
+ chattr
33
+ check-new-release
34
+ chmod
35
+ chown
36
+ chrome-headless-shell
37
+ claude
38
+ claude.exe
39
+ cleanup
40
+ clear
41
+ clear_console
42
+ cloudcode_cli
43
+ cmp
44
+ cnf-update-db
45
+ code
46
+ code-server
47
+ collect2
48
+ command-not-found
49
+ containerd
50
+ cp
51
+ cron
52
+ crontab
53
+ curl
54
+ cut
55
+ dash
56
+ date
57
+ dbus-daemon
58
+ dbus-send
59
+ dbus.postinst
60
+ deb-systemd-invoke
61
+ debian-sa1
62
+ diff
63
+ dircolors
64
+ dirname
65
+ dmidecode
66
+ docker
67
+ dpkg
68
+ dpkg-db-backup
69
+ dpkg-deb
70
+ dpkg-maintscript-helper
71
+ dpkg-preconfigure
72
+ dpkg-query
73
+ dpkg-split
74
+ dpkg-statoverride
75
+ du
76
+ echo
77
+ ed
78
+ env
79
+ envsubst
80
+ expr
81
+ file
82
+ find
83
+ findmnt
84
+ flock
85
+ friendly-recovery
86
+ g++
87
+ gcc
88
+ gdbus
89
+ gemini
90
+ getconf
91
+ getent
92
+ getopt
93
+ gettext
94
+ gh
95
+ git
96
+ git-remote-https
97
+ git-submodule
98
+ gpg-connect-agent
99
+ gpgconf
100
+ gpgv
101
+ grep
102
+ groups
103
+ gtk-update-icon-cache
104
+ gzip
105
+ head
106
+ helpztags
107
+ hicolor-icon-theme.postinst
108
+ hostname
109
+ htop
110
+ http
111
+ https
112
+ id
113
+ install
114
+ ip
115
+ jq
116
+ kmod
117
+ ld
118
+ ld-linux-x86-64.so.2
119
+ ldconfig
120
+ ldconfig.real
121
+ ldd
122
+ less
123
+ lesspipe
124
+ libc-bin.postinst
125
+ ln
126
+ local
127
+ locale
128
+ locale-check
129
+ login
130
+ logrotate
131
+ ls
132
+ lsb_release
133
+ lsmod
134
+ lspci
135
+ lxc-start
136
+ lxc-start (deleted)
137
+ make
138
+ man-db
139
+ man-db.postinst
140
+ mandb
141
+ master
142
+ memfd:lxc-attach (deleted)
143
+ mkdir
144
+ mktemp
145
+ mmdc
146
+ mount
147
+ mv
148
+ nano
149
+ netplan
150
+ node
151
+ node-gyp
152
+ npm
153
+ nvidia-modprobe
154
+ nvidia-smi
155
+ packagekitd
156
+ pager
157
+ pdfinfo
158
+ pdflatex
159
+ pet
160
+ pgrep
161
+ pickup
162
+ ping
163
+ pip
164
+ pkill
165
+ pkttyagent
166
+ polkit-agent-helper-1
167
+ polkitd
168
+ polkitd.postinst
169
+ polkitd.postrm
170
+ polkitd.prerm
171
+ postconf
172
+ postdrop
173
+ postfix-instance-generator
174
+ prebuild-install
175
+ preinst
176
+ printenv
177
+ printf
178
+ ps
179
+ pwsh
180
+ pyenv
181
+ pyenv-commands
182
+ pyenv-exec
183
+ pyenv-hooks
184
+ pyenv-init
185
+ pyenv-latest
186
+ pyenv-prefix
187
+ pyenv-rehash
188
+ pyenv-sh-activate
189
+ pyenv-version-file
190
+ pyenv-version-file-read
191
+ pyenv-version-name
192
+ pyenv-versions
193
+ pyenv-virtualenv-init
194
+ pyenv-virtualenv-prefix
195
+ pyenv-which
196
+ python
197
+ python3
198
+ python3.12
199
+ qmgr
200
+ readlink
201
+ realpath
202
+ release-upgrade-motd
203
+ reset
204
+ rg
205
+ rm
206
+ rmdir
207
+ rsyslogd
208
+ ruff
209
+ run-parts
210
+ sa1
211
+ sa2
212
+ sadc
213
+ sar.sysstat
214
+ savelog
215
+ scp
216
+ sed
217
+ sendmail
218
+ service
219
+ setcap
220
+ setpriv
221
+ sgml-base.postinst
222
+ sh
223
+ sha256sum
224
+ shred
225
+ sleep
226
+ snap
227
+ snapd
228
+ snapd-env-generator
229
+ snapd-generator
230
+ sort
231
+ ssh
232
+ sshd
233
+ sshd-socket-generator
234
+ stat
235
+ store
236
+ stty
237
+ sudo
238
+ sysstat
239
+ systemctl
240
+ systemd
241
+ systemd-cryptsetup-generator
242
+ systemd-debug-generator
243
+ systemd-executor
244
+ systemd-executor (deleted)
245
+ systemd-fstab-generator
246
+ systemd-getty-generator
247
+ systemd-hibernate-resume-generator
248
+ systemd-integritysetup-generator
249
+ systemd-journald
250
+ systemd-logind
251
+ systemd-logind (deleted)
252
+ systemd-networkd
253
+ systemd-networkd-wait-online
254
+ systemd-rc-local-generator
255
+ systemd-resolved
256
+ systemd-run-generator
257
+ systemd-system-update-generator
258
+ systemd-sysusers
259
+ systemd-sysv-generator
260
+ systemd-timedated
261
+ systemd-tmpfiles
262
+ systemd-tty-ask-password-agent
263
+ systemd-user-runtime-dir
264
+ systemd-veritysetup-generator
265
+ systemd-xdg-autostart-generator
266
+ tail
267
+ tar
268
+ tee
269
+ test
270
+ timeout
271
+ tmux
272
+ touch
273
+ tr
274
+ trivial-rewrite
275
+ tty
276
+ ubuntu-distro-info
277
+ uname
278
+ unattended-upgrade
279
+ uniq
280
+ unzip
281
+ update-alternatives
282
+ update-catalog
283
+ update-icon-caches
284
+ update-motd-unattended-upgrades
285
+ update-xmlcatalog
286
+ uptime
287
+ vim
288
+ vim-common.postinst
289
+ vim-common.postrm
290
+ vim-runtime.postinst
291
+ vim-runtime.postrm
292
+ vim-tiny.postinst
293
+ vim-tiny.prerm
294
+ vim.postinst
295
+ vim.prerm
296
+ visudo
297
+ vsce-sign
298
+ wc
299
+ wget
300
+ which
301
+ whoami
302
+ whois
303
+ xargs
304
+ xdpyinfo
305
+ xrandr
306
+ xxd
307
+ zcat
308
+ zgrep
reproduction/art-campaign/ARTcampaign.md ADDED
@@ -0,0 +1,383 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # ART campaign grounding — 2026-04-19 late evening
2
+
3
+ (Prior content of this file — the v0.1 code-debt / experiment-protocol plan
4
+ for issues #23/#27/#32/#33/#34/#40 — was kicked off to 35-training-eval-harness
5
+ chat and is tracked in git + memory `project_v01_eod_2026_04_18.md`. This
6
+ file is being repurposed for a new, unrelated task in the same chat.)
7
+
8
+ ## Context
9
+
10
+ `run_atomic_trial.sh` just landed in the ART chat; dry-run against
11
+ `Invoke-AtomicTest <tech> -ShowDetailsBrief` produced
12
+ `docs/atr/linux-attack-verification.txt` showing which of the 12
13
+ v0.1-short-list techniques have atomics at the pinned SHA
14
+ `e575444941e27787bdb1b6ab881628fc2cc904b9`. Before executing Phase 3
15
+ (the 36-trial campaign that feeds `data/labels.csv`) user wants to
16
+ verify the attack strategy is:
17
+
18
+ - **Well thought out** — 12 techniques are the right 12 for v0.1.
19
+ - **Real-world practical** — atomics exercised are the realistic
20
+ variants, not trivial ones.
21
+ - **Demo-thesis aligned** — supply-chain behavioral chain story
22
+ (install → shell → recon → cred → exfil → persistence) from
23
+ `docs/vision/5-demo_brief.md`.
24
+
25
+ Everything that follows is read-only research + open questions. No code
26
+ changes until user approves via ExitPlanMode.
27
+
28
+ ---
29
+
30
+ ## Verified current state
31
+
32
+ **Wrapper (`mamba-edge/scripts/run_atomic_trial.sh`)**:
33
+ - Marker pattern: `touch /tmp/art_start_<TECH>_<TRIAL>` → execute → `touch /tmp/art_end_<TECH>_<TRIAL>`, each with immediate `rm -f`.
34
+ - `TEST_NUM` defaults to **1** (line 33) — no per-technique curation of *which* atomic under a technique gets run.
35
+ - Runs `pwsh -Command "Invoke-AtomicTest ${TECH} -TestNumbers ${TEST_NUM} -Execute -ExecutionLogPath ${LOG_PATH}"` then `-Cleanup`.
36
+ - Regex-validated technique input: `[A-Z0-9.]+`.
37
+ - Preserves `-Execute` RC; `-Cleanup` failures don't overwrite it.
38
+
39
+ **Label generator (`mamba-edge/scripts/generate_labels_from_markers.py:63` — `MITRE_MAP`)**, 12 techniques:
40
+ ```
41
+ T1059.004 Unix Shell execution
42
+ T1053.003 Cron persistence
43
+ T1548.003 Sudo and Sudo Caching privilege_escalation
44
+ T1003.008 /etc/passwd and /etc/shadow credential_access
45
+ T1057 Process Discovery discovery
46
+ T1082 System Information Discovery discovery
47
+ T1105 Ingress Tool Transfer command_and_control
48
+ T1070.004 File Deletion defense_evasion
49
+ T1222.002 Linux/Mac File/Dir Perms Mod defense_evasion
50
+ T1055.008 Ptrace System Calls privilege_escalation ← BROKEN
51
+ T1027 Obfuscated Files or Information defense_evasion
52
+ T1033 System Owner/User Discovery discovery
53
+ ```
54
+
55
+ **Atomic availability at SHA `e575444` (from `linux-attack-verification.txt`)**:
56
+ - 11 of 12 techniques have atomics: counts range 1 (T1057/T1027/T1033) up to 17 (T1059.004), 14 (T1222.002), 9 (T1082/T1105).
57
+ - **T1055.008 is MISSING** at this SHA: `atomics/T1055.008/T1055.008.yaml does not exist`. The wrapper will fail/no-op for this technique.
58
+ - T1082 starts at `-3` (atomics `-1`/`-2` are non-Linux); wrapper's default `TEST_NUM=1` would error on T1082.
59
+
60
+ **Tetragon policy coverage (`mamba-edge/config/tetragon/*.yaml`)** — 6 policies deployed:
61
+ - `file-monitor.yaml` (fd_install, do_unlinkat)
62
+ - `privilege-monitor.yaml` (commit_creds → setuid/sudo/capset)
63
+ - `network-monitor.yaml` (tcp_connect, tcp_close, inet_csk_accept, udp_sendmsg:53)
64
+ - `memory-monitor.yaml`, `injection-monitor.yaml`, `attribute-monitor.yaml`
65
+
66
+ **Report doc (`docs/releases/v0.1-course-milestone/v0.1-attack-coverage.md` v1.2)** defines a **DIFFERENT** 12-technique list:
67
+ ```
68
+ T1059.004, T1059.006 Python, T1053.003, T1543.002 Systemd,
69
+ T1546.004 .bashrc, T1548.001 Setuid, T1548.003, T1070.004,
70
+ T1222.002, T1003.008, T1552.001 Creds in Files, T1057
71
+ ```
72
+ Overlap with MITRE_MAP: 7 techniques (T1059.004, T1053.003, T1548.003, T1003.008, T1057, T1070.004, T1222.002).
73
+ Drift (MITRE_MAP has these, doc doesn't): T1082, T1105, T1055.008, T1027, T1033 (**3 are discovery; 1 is broken**).
74
+ Drift (doc has these, MITRE_MAP doesn't): T1059.006, T1543.002, T1546.004, T1548.001, T1552.001 (**demo-aligned persistence + priv-esc + cred access**).
75
+
76
+ **Demo brief (`docs/vision/5-demo_brief.md`)** — V1 narrative is a **supply-chain behavioral chain**: package install → unexpected shell/interpreter → suspicious child lineage → sensitive file read → outbound connection → persistence.
77
+
78
+ **Track A (synthetic npm payload)** — `v0.1-attack-coverage.md` §3 describes this as the *centerpiece* (8 trials, `T1195.002`, `fake-malicious-npm/` with fake-listener + fake-secrets + postinstall.js). **Not represented** in MITRE_MAP or in `run_atomic_trial.sh`; descoped from current code.
79
+
80
+ ---
81
+
82
+ ## Gaps identified
83
+
84
+ 1. **T1055.008 is broken at pinned SHA.** The wrapper will fail on this technique; labels pipeline will not produce intervals for it. Needs resolution (drop or swap) before the campaign runs.
85
+
86
+ 2. **MITRE_MAP vs attack-coverage.md v1.2 drift.** The code and the report doc disagree on 5 techniques each direction. Code was updated after the doc (per `project_art_track_b_status.md`, 12-list confirmed by chat A on 2026-04-18), but the doc-side list is more demo-aligned (richer persistence and priv-esc coverage; no broken technique).
87
+
88
+ 3. **Discovery over-representation in MITRE_MAP.** 3 of 12 techniques are discovery (T1057 Process Discovery, T1082 System Info, T1033 Owner/User). Discovery is noisy in baseline — ps/whoami/uname happen constantly — so these atomics produce weak discriminator signal. The attack-coverage.md list has only 1 discovery (T1057) and reallocates the slots to persistence and cred access.
89
+
90
+ 4. **Tactic balance skew.** MITRE_MAP has 1 execution, 1 persistence vs doc's 2 + 3. Persistence is central to the supply-chain thesis (the "postinstall drops a cronjob" stage) — underrepresenting it weakens the demo narrative mapping.
91
+
92
+ 5. **`TEST_NUM=1` default without per-technique curation.** Some `-1` atomics are trivial (T1548.003-1 is just `sudo`; T1070.004-1 is `rm` on a test file; T1222.002-1 is a single `chmod`); others are reasonable (T1059.004-1 creates + runs a bash script; T1003.008-1 reads `/etc/shadow`; T1027-1 decodes base64). T1082-1 doesn't exist on Linux (first is `-3`). Running `TEST_NUM=1` for everything means some trials are barely distinguishable from baseline.
93
+
94
+ 6. **Three trials per technique currently run the SAME atomic.** Track X campaign spec is "12 techniques × 3 trials = 36 trials", but if all 3 trials use the same atomic number, you get 3 near-identical event sequences instead of 3 variations of the technique. Tetragon techniques like T1059.004 have 17 atomics covering substantially different behavior (basic shell, LinEnum, SUID harvest, awk shell, cpan shell, busybox shell, obfuscated cmdline). Using a different atomic number per trial would give real variation.
95
+
96
+ 7. **No chain structure anywhere in v0.1 execution.** The demo brief thesis is a **chain** (install → shell → recon → cred → exfil → persistence). Running 36 independent atomics inside 36 separate marker pairs doesn't produce the chain narrative. Track A was the mechanism for this (synthetic npm payload runs all 6 stages in one trial) — and it's descoped from code.
97
+
98
+ ---
99
+
100
+ ## Recommended approach
101
+
102
+ Below are my defaults; user decisions drive the final plan.
103
+
104
+ ### A. Replace T1055.008
105
+ **Default: swap to T1546.004 (Unix Shell Configuration Modification — `.bashrc` / `.profile` append).**
106
+ Rationale: (a) it's in `attack-coverage.md` already so no doc drift; (b) fits the supply-chain persistence stage; (c) Tetragon `fd_install` + `process_exec` cover it cleanly (bash writes to `~/.bashrc`, shell spawn reads it on next session); (d) need to verify `T1546.004.yaml` exists at pinned SHA in the pre-campaign pass.
107
+
108
+ ### B. Technique list source of truth
109
+ **Default: reconcile toward `attack-coverage.md` v1.2 list** (12 techniques: drop T1082, T1105, T1055.008, T1027, T1033; add T1059.006, T1543.002, T1546.004, T1548.001, T1552.001).
110
+ Rationale: the doc list is demo-aligned (3 persistence, 2 cred-access, 2 priv-esc, 2 execution — proper tactic balance for supply-chain chain), and dropping the 3 discovery techniques removes the baseline-signal-noise problem. Code change is localized to `MITRE_MAP` in `generate_labels_from_markers.py:63`.
111
+ **Alternative:** keep MITRE_MAP as-is (only swap T1055.008 → something), update attack-coverage.md in a follow-up doc-sync PR. Lighter touch but keeps discovery overweighting and broken `T1082 TEST_NUM=1`.
112
+
113
+ ### C. Atomic number curation
114
+ **Default: run 3 different atomic numbers per technique (one per trial), selected for (i) Linux-validity, (ii) non-trivial behavior, (iii) coverage by deployed Tetragon policies.**
115
+ Produces a short curation table before the campaign runs. Takes ~45 min to build (read 12 × 3 = 36 atomic YAMLs in `~/AtomicRedTeam/atomics/`), spot-check each via `Invoke-AtomicTest -ShowDetails`, record chosen `-TestNumbers` in a new `mamba-edge/config/art_campaign.yaml` or equivalent. Wrapper already accepts `TEST_NUM` as 3rd positional arg — just need to pass different values per trial.
116
+
117
+ ### D. Capstone chain for demo narrative
118
+ **Default: add a single "capstone chain" trial at the end of the 36 independent trials.**
119
+ One marker pair (`/tmp/art_start_CAPSTONE_1` → `/tmp/art_end_CAPSTONE_1`), inside which 5 atomics run back-to-back in demo-brief order:
120
+ 1. T1059.004 shell spawn
121
+ 2. T1057 discovery
122
+ 3. T1003.008 shadow read
123
+ 4. T1105 curl outbound (loopback target — safe in lab)
124
+ 5. T1070.004 rm evidence
125
+
126
+ Label with a new pseudo-technique id added to MITRE_MAP (`CHAIN-CAP01` or similar, or label with the dominant technique and describe sequence in `notes`).
127
+ Keeps the marker/label pipeline unchanged. Total run time ~60s. Recovers the supply-chain narrative without restoring Track A's Node-payload engineering (fake-listener.py, fake-secrets, postinstall.js, crontab manipulation).
128
+
129
+ ### E. Track A status
130
+ **Default: formally defer to v0.2.** Document the deferral in `v0.1-attack-coverage.md` next to the capstone-chain description. Engineering cost (~300 LOC + 3 scripts + local HTTP listener + fake-secrets harness) is not worth it this close to the 2026-04-29 deadline. Capstone chain (D) is the v0.1 substitute.
131
+
132
+ ### F. Campaign order & safety
133
+ Unchanged from Track X spec:
134
+ - Pre-flight: `validate_labels.py` + `Invoke-AtomicTest -CheckPrereqs` for every curated atomic.
135
+ - Run from `primary_user` (uid 1000) per `feedback_art_attacks_preserve_persona.md`; let Invoke-AtomicTest `sudo` internally for root-needing atomics.
136
+ - No `sudo` wrap of the wrapper (contaminates persona attribution).
137
+ - ZFS snapshot + `pct snapshot` rollback points before Phase 3 (already taken per the 04-19T04:23 install row in the collection log).
138
+ - Pause baseline cron harness during the attack session; resume at the end.
139
+
140
+ ---
141
+
142
+ ## User decisions received (2026-04-19 late evening)
143
+
144
+ 1. **Technique list size**: "I would like to consider adding even more. I don't think 12 is enough. I would love to see more and what the mamba model can detect over the other models."
145
+ → Scope expands beyond 12; choose from 14 / 17 (recommended) / 20 in the follow-up AskUserQuestion.
146
+ 2. **Demo narrative**: "36 independent + 1 capstone chain AND track A (why did we leave track A behind?)"
147
+ → All three: independent trials + capstone chain + Track A restored.
148
+
149
+ ### Why Track A was deferred (answer to user's question)
150
+
151
+ On 2026-04-18 during v0.1 scope-lock (memory `project_v01_scope_lock_and_issue_audit.md`), engineering budget was the binding constraint and we pivoted to all-ART breadth for simpler engineering. This was a mistake in hindsight:
152
+
153
+ - Track A maps directly to `docs/vision/5-demo_brief.md`'s supply-chain thesis (it IS the scenario that doc was written around).
154
+ - Supply-chain attacks are the current highest-profile threat class (2024 `shai-hulud`-style npm worms, 2025 obfuscated postinstall campaigns).
155
+ - Track A gives the report a concrete, named, narrative centerpiece — "our model caught a synthetic npm supply-chain chain" reads stronger than "our model detected per-technique anomalies."
156
+ - Engineering cost (~300 LOC across 6 scripts + tiny HTTP listener + fake-secrets setup) is achievable in ~1.5 days.
157
+
158
+ **Restoration decision: full Track A per `v0.1-attack-coverage.md` §3.** 8 trials with the variation axes already defined in §3.4 (baseline chain, skip-persistence variant, DNS-exfil variant, reversed-order variant, etc.). No scope reductions.
159
+
160
+ ---
161
+
162
+ ## Proposed expanded technique list (user picks size)
163
+
164
+ Baseline is the `attack-coverage.md` v1.2 list (12 techniques), plus additions from the MITRE ATT&CK Linux space that are safe, lab-compatible, and produce distinct Tetragon signatures. All additions verified to exist in ART; final YAML spot-check happens in the pre-campaign pass.
165
+
166
+ ### Option X: 14 techniques (conservative)
167
+
168
+ attack-coverage.md 12 + 2 breadth additions from MITRE_MAP:
169
+
170
+ | Tactic | Techniques |
171
+ |---|---|
172
+ | Execution (2) | T1059.004 Unix Shell, T1059.006 Python |
173
+ | Persistence (3) | T1053.003 Cron, T1543.002 Systemd, T1546.004 .bashrc |
174
+ | Priv Esc (2) | T1548.001 Setuid, T1548.003 Sudo |
175
+ | Defense Evasion (2) | T1070.004 File Deletion, T1222.002 Perms Mod |
176
+ | Credential Access (2) | T1003.008 /etc/shadow, T1552.001 Creds in Files |
177
+ | Discovery (1) | T1057 Process Discovery |
178
+ | C2 / Ingress (2) | **T1082 System Info**, **T1105 Ingress Tool Transfer** |
179
+
180
+ 14 × 3 = 42 ART trials + 1 capstone + 8 Track A = **51 labeled intervals**.
181
+
182
+ ### Option Y: 17 techniques (recommended — well-balanced breadth)
183
+
184
+ Option X + 3 more tactic-reinforcing additions:
185
+
186
+ | Tactic | Techniques |
187
+ |---|---|
188
+ | Execution (2) | T1059.004 Unix Shell, T1059.006 Python |
189
+ | Persistence (3) | T1053.003 Cron, T1543.002 Systemd, T1546.004 .bashrc |
190
+ | Priv Esc (3) | T1548.001 Setuid, T1548.003 Sudo, **T1574.006 LD_PRELOAD** |
191
+ | Defense Evasion (3) | T1070.004 File Deletion, T1222.002 Perms Mod, **T1027 Obfuscated** |
192
+ | Credential Access (2) | T1003.008 /etc/shadow, T1552.001 Creds in Files |
193
+ | Discovery (2) | T1057 Process Discovery, T1082 System Info |
194
+ | C2 / Ingress (2) | T1105 Ingress Tool Transfer, **T1562.001 Impair Defenses** |
195
+
196
+ 17 × 3 = 51 ART trials + 1 capstone + 8 Track A = **60 labeled intervals**.
197
+
198
+ Rationale for the +3 over Option X:
199
+ - **T1574.006 LD_PRELOAD** — sophisticated priv-esc via shared-library injection; distinct Tetragon `fd_install` + `process_exec` signature with `LD_PRELOAD=` in env.
200
+ - **T1027 Obfuscated** — common evasion pattern (`echo <base64> | base64 -d | bash`); pairs nicely with the encoder's `cmdline_max_entropy` + `cmdline_compression_ratio` features (#23).
201
+ - **T1562.001 Impair Defenses** — disabling/modifying logging or monitoring tools; important adversary tradecraft for the report's Discussion section.
202
+
203
+ ### Option Z: 20 techniques (aggressive — comprehensive)
204
+
205
+ Option Y + 3 more:
206
+
207
+ | Tactic | Techniques |
208
+ |---|---|
209
+ | Execution (3) | T1059.004 Unix Shell, T1059.006 Python, **T1569.002 Service Execution** |
210
+ | Persistence (4) | T1053.003 Cron, T1543.002 Systemd, T1546.004 .bashrc, **T1098.004 SSH Authorized Keys** |
211
+ | Priv Esc (3) | T1548.001 Setuid, T1548.003 Sudo, T1574.006 LD_PRELOAD |
212
+ | Defense Evasion (4) | T1070.004 File Deletion, T1222.002 Perms Mod, T1027 Obfuscated, **T1036.005 Masquerading** |
213
+ | Credential Access (2) | T1003.008 /etc/shadow, T1552.001 Creds in Files |
214
+ | Discovery (2) | T1057 Process Discovery, T1082 System Info |
215
+ | C2 / Ingress (2) | T1105 Ingress Tool Transfer, T1562.001 Impair Defenses |
216
+
217
+ 20 × 3 = 60 ART trials + 1 capstone + 8 Track A = **69 labeled intervals**.
218
+
219
+ ### Trial count feasibility
220
+
221
+ Per `v0.1-attack-coverage.md` §6.2, the attack session is planned as "one sitting, ~3–4 hours". Each trial is ~30–60s attack + ~10–30s cleanup + ~5–10s between-trial bookkeeping = ~60–100s end-to-end. Reality:
222
+
223
+ | Option | Trials | Estimated session length |
224
+ |---|---|---|
225
+ | X (51) | 42 + 1 + 8 | ~60–85 min attack + setup |
226
+ | Y (60) | 51 + 1 + 8 | ~70–100 min |
227
+ | Z (69) | 60 + 1 + 8 | ~85–115 min |
228
+
229
+ All three fit the 3–4 hour session budget with headroom. The binding constraint isn't attack time — it's atomic-curation prep time (reading atomic YAMLs + spot-checking prereqs):
230
+
231
+ | Option | Atomic curation time | Code touch for MITRE_MAP + tests |
232
+ |---|---|---|
233
+ | X | ~45 min (14 × 3 = 42 atomics) | small |
234
+ | Y | ~60 min (17 × 3 = 51 atomics) | small |
235
+ | Z | ~75 min (20 × 3 = 60 atomics) | small |
236
+
237
+ ---
238
+
239
+ ## Revised execution order (post-approval)
240
+
241
+ 1. Apply final technique list (edit `MITRE_MAP` in `generate_labels_from_markers.py:63` + update `test_generate_labels_from_markers.py` fixtures).
242
+ 2. Atomic curation pass: read atomic YAMLs, build `mamba-edge/config/art_campaign.yaml` with 3 curated `TEST_NUM`s per technique.
243
+ 3. Wrapper extension: `run_atomic_trial.sh` already accepts TEST_NUM as arg 3; add `run_atomic_chain.sh` for the capstone (5 atomics in one marker pair).
244
+ 4. **Track A engineering** (~1.5 days):
245
+ - `mamba-edge/attacks/fake-malicious-npm/` package (`package.json`, `index.js`, `scripts/postinstall.js`, `scripts/stage1-recon.sh`, `scripts/stage2-exfil.sh`, `scripts/stage3-persistence.sh`, `scripts/cleanup.sh`).
246
+ - `mamba-edge/attacks/fake-malicious-npm/test-harness/` (`fake-listener.py`, `setup-fake-secrets.sh`, `run-trial.sh`).
247
+ - Add `T1195.002` → `_TechniqueMeta("Compromise Software Supply Chain", "initial_access")` to MITRE_MAP.
248
+ - README.md with prominent "RESEARCH ARTIFACT" banner.
249
+ 5. Doc sync: `v0.1-attack-coverage.md` v1.2 → v1.3: new N-technique list + Track A restoration note (rescind the "descoped" implication) + capstone chain section.
250
+ 6. Commit sequence (Conventional Commits): `feat(art)` for MITRE_MAP + curation + chain, `feat(attacks)` for Track A payload, `docs(art)` for attack-coverage v1.3.
251
+ 7. Pre-campaign verification: `Invoke-AtomicTest <tech> -ShowDetails -TestNumbers N` for every curated atomic; `Invoke-AtomicTest <tech> -CheckPrereqs`; snapshot output to `docs/atr/linux-attack-verification-curated.txt`.
252
+ 8. Campaign execution is a separate session.
253
+
254
+ ---
255
+
256
+ ## Final decisions (all locked)
257
+
258
+ - **Technique list**: Option Z — 20 techniques (see table below). 60 ART trials + 1 capstone + 8 Track A = **69 labeled intervals**.
259
+ - **Demo narrative**: 36+ independent + 1 capstone chain + full Track A.
260
+ - **Atomic curation**: 3 different atomic numbers per technique (total 60 curated atomics). Recorded in new `mamba-edge/config/art_campaign.yaml`.
261
+ - **Track A**: restored fully per `v0.1-attack-coverage.md` §3 — 8 trials, `T1195.002`, `mamba-edge/attacks/fake-malicious-npm/`. No scope reductions.
262
+
263
+ ### Final 20-technique list (Option Z)
264
+
265
+ | Tactic | Techniques (20) |
266
+ |---|---|
267
+ | Execution (3) | T1059.004 Unix Shell, T1059.006 Python, T1569.002 Service Execution |
268
+ | Persistence (4) | T1053.003 Cron, T1543.002 Systemd, T1546.004 .bashrc, T1098.004 SSH Authorized Keys |
269
+ | Priv Esc (3) | T1548.001 Setuid, T1548.003 Sudo, T1574.006 LD_PRELOAD |
270
+ | Defense Evasion (4) | T1070.004 File Deletion, T1222.002 Perms Mod, T1027 Obfuscated, T1036.005 Masquerading |
271
+ | Credential Access (2) | T1003.008 /etc/shadow, T1552.001 Creds in Files |
272
+ | Discovery (2) | T1057 Process Discovery, T1082 System Info |
273
+ | C2 / Ingress (2) | T1105 Ingress Tool Transfer, T1562.001 Impair Defenses |
274
+
275
+ Plus the Track A pseudo-technique `T1195.002 Compromise Software Supply Chain` (initial_access) and the capstone chain (labeled with a new `CHAIN-CAP01` pseudo-id or with `T1105` dominant + `notes` description; decide at commit time).
276
+
277
+ ### Capstone chain composition (5 atomics, one marker pair)
278
+
279
+ In demo-brief order, each step is an atomic from the 20-technique list above:
280
+
281
+ 1. `T1059.004` — Unix shell spawn
282
+ 2. `T1057` — Process discovery
283
+ 3. `T1003.008` — /etc/shadow read
284
+ 4. `T1105` — curl loopback (ingress tool transfer, safe lab target `127.0.0.1`)
285
+ 5. `T1070.004` — rm evidence
286
+
287
+ Runtime ~60s inside one `/tmp/art_start_CAPSTONE_1` → `/tmp/art_end_CAPSTONE_1` marker pair.
288
+
289
+ ### Final labeled-interval tally
290
+
291
+ - 20 techniques × 3 trials (3 different atomic numbers each) = **60 independent intervals**
292
+ - 1 capstone chain = **1 interval**
293
+ - 8 Track A trials = **8 intervals**
294
+ - **Total: 69 labeled attack intervals**
295
+
296
+ This is the v0.1 evaluation set `data/labels.csv` will contain at Phase 3 close.
297
+
298
+ ---
299
+
300
+ ## Final execution order (post-approval)
301
+
302
+ 1. **MITRE_MAP update** (`mamba-edge/scripts/generate_labels_from_markers.py:63`): replace the 12-entry dict with the 20-technique list + `T1195.002` for Track A + optional `CHAIN-CAP01` for capstone. Update `mamba-edge/tests/test_generate_labels_from_markers.py` fixtures.
303
+ 2. **Atomic curation** (~75 min): read 60 atomic YAMLs (20 × 3) under `~/AtomicRedTeam/atomics/T*/T*.yaml`; spot-check each via `Invoke-AtomicTest <tech> -ShowDetails -TestNumbers N`; record `{technique: [test_num_trial1, test_num_trial2, test_num_trial3]}` in new `mamba-edge/config/art_campaign.yaml`.
304
+ 3. **Capstone wrapper**: new `mamba-edge/scripts/run_atomic_chain.sh` wrapping 5 atomics inside one marker pair. Reuses `run_atomic_trial.sh`'s marker logic but with a single wrapping `art_start_CAPSTONE_1` / `art_end_CAPSTONE_1` pair around the five `Invoke-AtomicTest` calls.
305
+ 4. **Track A engineering** (~1.5 days, full per `v0.1-attack-coverage.md` §3):
306
+ - `mamba-edge/attacks/fake-malicious-npm/package.json` — postinstall hook.
307
+ - `mamba-edge/attacks/fake-malicious-npm/index.js` — benign front door.
308
+ - `mamba-edge/attacks/fake-malicious-npm/scripts/postinstall.js` — Node → shell entry point.
309
+ - `mamba-edge/attacks/fake-malicious-npm/scripts/stage1-recon.sh` — reads fake creds.
310
+ - `mamba-edge/attacks/fake-malicious-npm/scripts/stage2-exfil.sh` — base64 + `curl` loopback POST.
311
+ - `mamba-edge/attacks/fake-malicious-npm/scripts/stage3-persistence.sh` — crontab entry (cleaned up).
312
+ - `mamba-edge/attacks/fake-malicious-npm/scripts/cleanup.sh`.
313
+ - `mamba-edge/attacks/fake-malicious-npm/test-harness/fake-listener.py` — 127.0.0.1:8888 silent POST sink.
314
+ - `mamba-edge/attacks/fake-malicious-npm/test-harness/setup-fake-secrets.sh` — `~/.fake_secrets/*.env` with obvious dummy content.
315
+ - `mamba-edge/attacks/fake-malicious-npm/test-harness/run-trial.sh` — orchestrator that writes the `labels.csv` row (or emits markers for the labeler to pick up).
316
+ - `mamba-edge/attacks/fake-malicious-npm/README.md` — RESEARCH ARTIFACT banner.
317
+ 5. **Doc sync**: `v0.1-attack-coverage.md` v1.2 → v1.3. Rescind "descoped" framing of Track A, update §5 with the 20-technique list (replacing v1.2's 12-list), add capstone-chain section, document the per-technique 3-different-atomic approach, preserve the safety checklist in §10.
318
+ 6. **Commits** (Conventional, on `release/v0.1`, no push):
319
+ - `feat(art): expand Track B to 20 techniques with curated atomics`
320
+ - `feat(art): add capstone chain wrapper and pseudo-technique label`
321
+ - `feat(attacks): restore synthetic npm supply-chain payload (Track A)`
322
+ - `docs(releases/v0.1): attack coverage v1.3 — 20 techniques + Track A + capstone`
323
+ 7. **Pre-campaign verification**: run `-CheckPrereqs` on all 60 curated atomics; snapshot output to `docs/atr/linux-attack-verification-curated.txt`; dry-run `fake-malicious-npm` end-to-end on scratch dir and confirm fake-listener receives POST.
324
+ 8. **Campaign execution** (Phase 3, separate hands-on session): Track A 8 trials first, then 60 independent ART trials (tactic-ordered), then 1 capstone chain.
325
+
326
+ ---
327
+
328
+ ## Verification
329
+
330
+ - `grep -n 'T1569.002\\|T1098.004\\|T1574.006\\|T1036.005\\|T1562.001\\|T1195.002' mamba-edge/scripts/generate_labels_from_markers.py` — all 6 new techniques + Track A id present.
331
+ - `pytest mamba-edge/tests/test_generate_labels_from_markers.py` → green.
332
+ - `docs/atr/linux-attack-verification-curated.txt` shows 60 atomics × `-ShowDetails` plus 8 Track A trials × `-CheckPrereqs` dry-run.
333
+ - `fake-malicious-npm/` end-to-end dry-run leaves no artifacts (`~/.fake_secrets/` cleaned, crontab restored).
334
+ - `docs/releases/v0.1-course-milestone/v0.1-attack-coverage.md` §5 technique table matches `MITRE_MAP`; v1.3 change log entry present.
335
+ - `CLAUDE-activeContext.md` updated: "ART Phase 3 ready — 60 independent + 1 capstone + 8 Track A = 69 labeled intervals; curated atomics in `config/art_campaign.yaml`".
336
+ - `gh issue #30` comment pinned with the final list + estimated session runtime.
337
+
338
+ ---
339
+
340
+ ## Critical files (read-only references)
341
+
342
+ - `mamba-edge/scripts/run_atomic_trial.sh` — wrapper (77 lines; no changes needed unless Decision 3 is "capstone chain" — then add a wrapper variant for multi-atomic chains)
343
+ - `mamba-edge/scripts/generate_labels_from_markers.py:63` — `MITRE_MAP` (likely edited; tests in `mamba-edge/tests/test_generate_labels_from_markers.py` may need updates)
344
+ - `docs/releases/v0.1-course-milestone/v0.1-attack-coverage.md` — doc update required regardless (v1.2 → v1.3, reflecting finalized list + capstone chain + Track A deferral note)
345
+ - `docs/atr/linux-attack-verification.txt` — authoritative atomics-availability snapshot; do not edit, reference only
346
+ - `mamba-edge/config/tetragon/{file,privilege,network}-monitor.yaml` — confirm each curated atomic's syscalls are covered (read-only check)
347
+ - `~/AtomicRedTeam/atomics/T<tech>/T<tech>.yaml` — the atomic YAML files (read-only; source for curation table)
348
+ - `docs/vision/5-demo_brief.md` — chain thesis (no edits; just the anchor we're aligning to)
349
+ - `CLAUDE-activeContext.md` — update phase pointer + campaign-readiness state after decisions land
350
+
351
+ ---
352
+
353
+ ## Execution order (post-approval)
354
+
355
+ 1. **Apply Decision 1** (T1055.008 fate) and **Decision 2** (list reconciliation) — edit `MITRE_MAP` in `generate_labels_from_markers.py:63`, update tests.
356
+ 2. **Apply Decision 3** (demo narrative) — if capstone chain: add marker-chain logic to a new `mamba-edge/scripts/run_atomic_chain.sh` or extend existing wrapper; add CHAIN-CAP01 (or chosen id) to MITRE_MAP.
357
+ 3. **Atomic curation pass** — read the 12 × 3 atomic YAMLs, build a curation table at `mamba-edge/config/art_campaign.yaml` listing `{technique: [test_num_trial1, test_num_trial2, test_num_trial3]}`.
358
+ 4. **Doc sync** — update `v0.1-attack-coverage.md` v1.2 → v1.3: new 12-list, capstone-chain section, Track A deferral note.
359
+ 5. **Commit** — conventional commits, `feat(art): ...` scoped per change (list reconciliation + capstone + curation = 3 commits).
360
+ 6. **Pre-campaign verification** — run `Invoke-AtomicTest <tech> -ShowDetails -TestNumbers N` for each curated atomic; confirm atomic YAML exists and the executor is Linux-compatible.
361
+ 7. **Commit** the pre-flight spot-check output to `docs/atr/linux-attack-verification-curated.txt`.
362
+ 8. Campaign execution (Phase 3) is a separate session.
363
+
364
+ ---
365
+
366
+ ## Verification (how we'll know it's right)
367
+
368
+ - `python -c "from mamba_edge.scripts.generate_labels_from_markers import MITRE_MAP; assert 'T1055.008' not in MITRE_MAP or <swap_tech> in MITRE_MAP"` → passes.
369
+ - `pytest mamba-edge/tests/test_generate_labels_from_markers.py` → green.
370
+ - `docs/releases/v0.1-course-milestone/v0.1-attack-coverage.md` §5 technique table matches `MITRE_MAP`.
371
+ - `docs/atr/linux-attack-verification-curated.txt` shows every curated atomic returns a valid `-ShowDetails` response.
372
+ - `gh issue #30` comment summarizes: resolved list + capstone + curation approach; campaign ready to execute.
373
+ - `CLAUDE-activeContext.md` updated: "ART Phase 3 ready — 36 independent trials + 1 capstone chain; curated atomics in `config/art_campaign.yaml`".
374
+
375
+ ---
376
+
377
+ ## Out of scope tonight
378
+
379
+ - Executing the 36-trial campaign (Phase 3 — separate hands-on session).
380
+ - Implementing Track A synthetic npm payload (deferred to v0.2 regardless of tonight's decisions).
381
+ - The orchestrator v0.1 code-debt stream (issues #23/#27/#32/#33/#34/#40 — owned by 35-training-eval-harness chat; status tracked in memory `project_v01_eod_2026_04_18.md`).
382
+ - Training runs (blocked on Phase 3 labels.csv).
383
+ - Baseline cron pause/resume scripts (assumed pre-existing per `v0.1-attack-coverage.md` §6.1; verify before Phase 3 launch, not tonight).
reproduction/art-campaign/labels.csv ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ host,technique_id,technique_name,tactic,start_ts,end_ts,trial,delivery,notes
2
+ sandbox_host,CHAIN.CAP01,Capstone Chain v0.1 (Supply-Chain Narrative),execution,2026-04-20T01:59:31.920547970Z,2026-04-20T01:59:45.019566446Z,1,art,
3
+ sandbox_host,T1003.008,/etc/passwd and /etc/shadow,credential_access,2026-04-20T00:27:19.544699604Z,2026-04-20T00:27:21.324864544Z,1,art,
4
+ sandbox_host,T1003.008,/etc/passwd and /etc/shadow,credential_access,2026-04-20T00:27:26.330849154Z,2026-04-20T00:27:28.103071204Z,2,art,
5
+ sandbox_host,T1003.008,/etc/passwd and /etc/shadow,credential_access,2026-04-20T00:27:33.109311865Z,2026-04-20T00:27:34.913462371Z,3,art,
6
+ sandbox_host,T1027,Obfuscated Files or Information,defense_evasion,2026-04-20T00:26:44.362314451Z,2026-04-20T00:26:46.260972910Z,1,art,
7
+ sandbox_host,T1036.005,Match Legitimate Name or Location,defense_evasion,2026-04-20T00:27:01.981472730Z,2026-04-20T00:27:03.743904270Z,1,art,
8
+ sandbox_host,T1053.003,Cron,persistence,2026-04-19T23:21:15.318086413Z,2026-04-19T23:21:17.129137323Z,1,art,
9
+ sandbox_host,T1053.003,Cron,persistence,2026-04-19T23:21:22.134835969Z,2026-04-19T23:21:23.928586108Z,2,art,
10
+ sandbox_host,T1053.003,Cron,persistence,2026-04-19T23:21:28.934815378Z,2026-04-19T23:21:30.715299868Z,3,art,
11
+ sandbox_host,T1057,Process Discovery,discovery,2026-04-20T00:30:24.402125645Z,2026-04-20T00:30:26.221544426Z,1,art,
12
+ sandbox_host,T1059.004,Unix Shell,execution,2026-04-19T23:18:01.681365242Z,2026-04-19T23:18:06.719474616Z,1,art,
13
+ sandbox_host,T1059.004,Unix Shell,execution,2026-04-19T23:18:11.725866322Z,2026-04-19T23:18:20.496957475Z,2,art,
14
+ sandbox_host,T1059.004,Unix Shell,execution,2026-04-19T23:18:25.503038168Z,2026-04-19T23:18:27.525715696Z,3,art,
15
+ sandbox_host,T1059.006,Python,execution,2026-04-19T23:18:43.341638054Z,2026-04-19T23:18:45.368727507Z,1,art,
16
+ sandbox_host,T1059.006,Python,execution,2026-04-19T23:18:50.375362308Z,2026-04-19T23:18:52.376624040Z,2,art,
17
+ sandbox_host,T1059.006,Python,execution,2026-04-19T23:18:57.383084874Z,2026-04-19T23:20:59.334084903Z,3,art,
18
+ sandbox_host,T1070.004,File Deletion,defense_evasion,2026-04-20T00:25:41.526420121Z,2026-04-20T00:25:43.370513123Z,1,art,
19
+ sandbox_host,T1070.004,File Deletion,defense_evasion,2026-04-20T00:25:48.376868194Z,2026-04-20T00:25:50.213997357Z,2,art,
20
+ sandbox_host,T1070.004,File Deletion,defense_evasion,2026-04-20T00:25:55.220340905Z,2026-04-20T00:25:57.134692866Z,3,art,
21
+ sandbox_host,T1082,System Information Discovery,discovery,2026-04-20T00:30:41.953754187Z,2026-04-20T00:30:44.050727668Z,1,art,
22
+ sandbox_host,T1082,System Information Discovery,discovery,2026-04-20T00:30:49.057145569Z,2026-04-20T00:30:51.145539391Z,2,art,
23
+ sandbox_host,T1082,System Information Discovery,discovery,2026-04-20T00:30:56.152341899Z,2026-04-20T00:30:58.262445401Z,3,art,
24
+ sandbox_host,T1083,File and Directory Discovery,discovery,2026-04-20T00:31:13.943771670Z,2026-04-20T00:31:15.858295030Z,1,art,
25
+ sandbox_host,T1083,File and Directory Discovery,discovery,2026-04-20T00:31:20.864889022Z,2026-04-20T00:31:30.438740733Z,2,art,
26
+ sandbox_host,T1083,File and Directory Discovery,discovery,2026-04-20T00:31:35.445243985Z,2026-04-20T00:31:37.269590947Z,3,art,
27
+ sandbox_host,T1098.004,SSH Authorized Keys,persistence,2026-04-20T00:14:12.358631741Z,2026-04-20T00:14:14.020111589Z,1,art,
28
+ sandbox_host,T1105,Ingress Tool Transfer,command_and_control,2026-04-20T00:31:52.969294229Z,2026-04-20T00:31:55.663769554Z,1,art,
29
+ sandbox_host,T1105,Ingress Tool Transfer,command_and_control,2026-04-20T00:32:00.671484771Z,2026-04-20T00:32:03.128654301Z,2,art,
30
+ sandbox_host,T1105,Ingress Tool Transfer,command_and_control,2026-04-20T00:32:08.135220672Z,2026-04-20T00:32:10.559928579Z,3,art,
31
+ sandbox_host,T1195.002,Compromise Software Supply Chain,initial_access,2026-04-19T22:45:43.962209623Z,2026-04-19T22:45:45.235700734Z,1,art,
32
+ sandbox_host,T1195.002,Compromise Software Supply Chain,initial_access,2026-04-19T22:45:51.463755615Z,2026-04-19T22:45:52.756753227Z,2,art,
33
+ sandbox_host,T1195.002,Compromise Software Supply Chain,initial_access,2026-04-19T22:45:58.980795684Z,2026-04-19T22:46:00.263649678Z,3,art,
34
+ sandbox_host,T1195.002,Compromise Software Supply Chain,initial_access,2026-04-19T22:46:06.486450544Z,2026-04-19T22:46:07.774254382Z,4,art,
35
+ sandbox_host,T1195.002,Compromise Software Supply Chain,initial_access,2026-04-19T22:46:13.991062337Z,2026-04-19T22:46:15.276347100Z,5,art,
36
+ sandbox_host,T1195.002,Compromise Software Supply Chain,initial_access,2026-04-19T22:46:21.493778083Z,2026-04-19T22:46:22.784321750Z,6,art,
37
+ sandbox_host,T1195.002,Compromise Software Supply Chain,initial_access,2026-04-19T22:46:29.005948563Z,2026-04-19T22:46:30.278480726Z,7,art,
38
+ sandbox_host,T1195.002,Compromise Software Supply Chain,initial_access,2026-04-19T22:46:36.495816216Z,2026-04-19T22:46:37.785140546Z,8,art,
39
+ sandbox_host,T1222.002,Linux and Mac File and Directory Permissions Modification,defense_evasion,2026-04-20T00:26:12.922543247Z,2026-04-20T00:26:14.846949090Z,1,art,
40
+ sandbox_host,T1222.002,Linux and Mac File and Directory Permissions Modification,defense_evasion,2026-04-20T00:26:19.853198637Z,2026-04-20T00:26:21.783431606Z,2,art,
41
+ sandbox_host,T1222.002,Linux and Mac File and Directory Permissions Modification,defense_evasion,2026-04-20T00:26:26.789544755Z,2026-04-20T00:26:28.721170208Z,3,art,
42
+ sandbox_host,T1543.002,Systemd Service,persistence,2026-04-19T23:50:59.765650410Z,2026-04-19T23:51:26.521462122Z,1,art,
43
+ sandbox_host,T1543.002,Systemd Service,persistence,2026-04-19T23:52:39.894720646Z,2026-04-19T23:54:17.576136538Z,2,art,
44
+ sandbox_host,T1543.002,Systemd Service,persistence,2026-04-19T23:54:22.582266884Z,2026-04-19T23:54:45.571424667Z,3,art,
45
+ sandbox_host,T1546.004,Unix Shell Configuration Modification,persistence,2026-04-19T23:55:01.568631159Z,2026-04-19T23:55:03.406917745Z,1,art,
46
+ sandbox_host,T1546.004,Unix Shell Configuration Modification,persistence,2026-04-19T23:55:08.413312998Z,2026-04-19T23:55:10.202837668Z,2,art,
47
+ sandbox_host,T1546.004,Unix Shell Configuration Modification,persistence,2026-04-19T23:55:15.209673036Z,2026-04-19T23:55:17.010232860Z,3,art,
48
+ sandbox_host,T1548.001,Setuid and Setgid,privilege_escalation,2026-04-20T00:14:29.982832656Z,2026-04-20T00:14:36.862804138Z,1,art,
49
+ sandbox_host,T1548.001,Setuid and Setgid,privilege_escalation,2026-04-20T00:14:41.869002067Z,2026-04-20T00:14:48.769179332Z,2,art,
50
+ sandbox_host,T1548.001,Setuid and Setgid,privilege_escalation,2026-04-20T00:14:53.775331564Z,2026-04-20T00:14:55.585905430Z,3,art,
51
+ sandbox_host,T1548.003,Sudo and Sudo Caching,privilege_escalation,2026-04-20T00:15:11.314148281Z,2026-04-20T00:17:13.146519169Z,1,art,
52
+ sandbox_host,T1548.003,Sudo and Sudo Caching,privilege_escalation,2026-04-20T00:17:18.152866405Z,2026-04-20T00:17:19.949403438Z,2,art,
53
+ sandbox_host,T1548.003,Sudo and Sudo Caching,privilege_escalation,2026-04-20T00:17:24.955045263Z,2026-04-20T00:17:30.215813570Z,3,art,
54
+ sandbox_host,T1552.001,Credentials In Files,credential_access,2026-04-20T00:27:50.629812895Z,2026-04-20T00:27:52.498058179Z,1,art,
55
+ sandbox_host,T1552.001,Credentials In Files,credential_access,2026-04-20T00:27:57.504584495Z,2026-04-20T00:29:59.576360961Z,2,art,
56
+ sandbox_host,T1552.001,Credentials In Files,credential_access,2026-04-20T00:30:04.583017711Z,2026-04-20T00:30:08.574646575Z,3,art,
57
+ sandbox_host,T1562.001,Disable or Modify Tools,defense_evasion,2026-04-20T00:32:26.379801715Z,2026-04-20T00:33:17.755734382Z,1,art,
58
+ sandbox_host,T1562.001,Disable or Modify Tools,defense_evasion,2026-04-20T00:33:22.761738237Z,2026-04-20T00:33:25.101641640Z,2,art,
59
+ sandbox_host,T1562.001,Disable or Modify Tools,defense_evasion,2026-04-20T00:33:30.108255228Z,2026-04-20T00:33:32.422829942Z,3,art,
60
+ sandbox_host,T1574.006,Dynamic Linker Hijacking,privilege_escalation,2026-04-20T00:19:24.001827850Z,2026-04-20T00:19:25.776246404Z,1,art,
61
+ sandbox_host,T1574.006,Dynamic Linker Hijacking,privilege_escalation,2026-04-20T00:19:30.782463207Z,2026-04-20T00:19:32.576378213Z,2,art,
62
+ sandbox_host,CHAIN.CAP01,Capstone Chain v0.1 (Supply-Chain Narrative),execution,2026-04-20T00:49:29.090978396Z,2026-04-20T00:49:37.045509190Z,2,art,broken 4-atomic chain: T1057:1 -> T1003.008:1 -> T1105:27 -> T1070.004:1; T1059.004 skipped due to invocation arg error (trial arg missing)
reproduction/feature_definitions.json ADDED
@@ -0,0 +1,821 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_meta": {
3
+ "schema_version": "1.1",
4
+ "source_modules": [
5
+ "mamba-edge/src/features/v0_2_features.py",
6
+ "mamba-edge/src/processing/v0_2_behavior_builder.py",
7
+ "mamba-edge/src/core/v0_2_event_encoder.py"
8
+ ],
9
+ "note": "Schema 1.1 reflects the 3a' encoder revisions (4 promoted categoricals — f_path_category / f_dst_ip_category / f_dst_port_category / f_object_category — plus hash right-sizing on f_proc_name_hash and f_proc_cwd_hash, cwd depth 2 -> 4) and the v0.2.1 deep-aliasing anonymizer (basename allowlist + custom_tool_NNN, salted-HMAC exec_ids). Integer-coded categorical/bucket columns map an integer to a label + semantic description. String-coded categorical columns (path_category, object_category, dst_ip_category, dst_port_category) map the string value directly. Hash columns list bucket count and the raw-parquet column the parser hashed."
10
+ },
11
+ "categorical_int_columns": {
12
+ "f_event_type": {
13
+ "0": {
14
+ "label": "process_exec",
15
+ "desc": "Process execution event (Tetragon process_exec)"
16
+ },
17
+ "1": {
18
+ "label": "process_exit",
19
+ "desc": "Process termination event"
20
+ },
21
+ "2": {
22
+ "label": "process_kprobe",
23
+ "desc": "Kernel-probe event hooked by Tetragon (file/net/cred/mem)"
24
+ },
25
+ "3": {
26
+ "label": "OOV",
27
+ "desc": "Unknown event type — should not appear; sentinel for parser drift"
28
+ }
29
+ },
30
+ "f_kprobe_function": {
31
+ "0": {
32
+ "label": "fd_install",
33
+ "desc": "File descriptor install (file open in the kernel)"
34
+ },
35
+ "1": {
36
+ "label": "security_mmap_file",
37
+ "desc": "LSM hook for mmap() of a file"
38
+ },
39
+ "2": {
40
+ "label": "commit_creds",
41
+ "desc": "Credential change (setuid/setgid/capset)"
42
+ },
43
+ "3": {
44
+ "label": "tcp_connect",
45
+ "desc": "Outbound TCP connect()"
46
+ },
47
+ "4": {
48
+ "label": "tcp_close",
49
+ "desc": "TCP socket close"
50
+ },
51
+ "5": {
52
+ "label": "udp_sendmsg",
53
+ "desc": "UDP send (port 53 = DNS resolves)"
54
+ },
55
+ "6": {
56
+ "label": "inet_csk_accept",
57
+ "desc": "Inbound TCP accept (server-side)"
58
+ },
59
+ "7": {
60
+ "label": "do_unlinkat",
61
+ "desc": "File unlink (delete)"
62
+ },
63
+ "8": {
64
+ "label": "chmod_common",
65
+ "desc": "File permission change"
66
+ },
67
+ "9": {
68
+ "label": "security_file_mprotect",
69
+ "desc": "LSM hook for mprotect() (memory permission change)"
70
+ },
71
+ "10": {
72
+ "label": "sys_ptrace",
73
+ "desc": "ptrace() syscall (process injection / debugger attach)"
74
+ },
75
+ "11": {
76
+ "label": "sys_process_vm_writev",
77
+ "desc": "process_vm_writev() (cross-process memory write)"
78
+ },
79
+ "12": {
80
+ "label": "NOT_KPROBE",
81
+ "desc": "Row is process_exec / process_exit, not a kprobe event"
82
+ },
83
+ "13": {
84
+ "label": "OOV",
85
+ "desc": "Kprobe function not in the v0.2 vocabulary"
86
+ }
87
+ },
88
+ "f_kprobe_policy": {
89
+ "0": {
90
+ "label": "file-monitor",
91
+ "desc": "Tracing policy for file-system kprobes"
92
+ },
93
+ "1": {
94
+ "label": "memory-monitor",
95
+ "desc": "Tracing policy for memory-related kprobes"
96
+ },
97
+ "2": {
98
+ "label": "privilege-monitor",
99
+ "desc": "Tracing policy for credential / privilege kprobes"
100
+ },
101
+ "3": {
102
+ "label": "network-monitor",
103
+ "desc": "Tracing policy for network kprobes"
104
+ },
105
+ "4": {
106
+ "label": "injection-monitor",
107
+ "desc": "Tracing policy for process-injection kprobes"
108
+ },
109
+ "5": {
110
+ "label": "attribute-monitor",
111
+ "desc": "Tracing policy for file-attribute kprobes"
112
+ },
113
+ "6": {
114
+ "label": "NONE",
115
+ "desc": "Row is not a kprobe (process_exec / process_exit)"
116
+ },
117
+ "7": {
118
+ "label": "OOV",
119
+ "desc": "Tracing policy not in the v0.2 vocabulary"
120
+ }
121
+ },
122
+ "f_kprobe_action": {
123
+ "0": {
124
+ "label": "KPROBE_ACTION_POST",
125
+ "desc": "Tetragon emitted the event without taking action"
126
+ },
127
+ "1": {
128
+ "label": "NONE",
129
+ "desc": "Row is not a kprobe"
130
+ },
131
+ "2": {
132
+ "label": "OOV",
133
+ "desc": "Action not seen in v0.2 (e.g. KILL/OVERRIDE in future Tetragon)"
134
+ }
135
+ },
136
+ "f_proc_uid_bucket": {
137
+ "0": {
138
+ "label": "ROOT",
139
+ "desc": "uid == 0 (container root)"
140
+ },
141
+ "1": {
142
+ "label": "RESEARCH_USER",
143
+ "desc": "uid == 1000 (primary research user in container)"
144
+ },
145
+ "2": {
146
+ "label": "SYSTEM_LOW",
147
+ "desc": "1 <= uid <= 99 (historical Unix system accounts)"
148
+ },
149
+ "3": {
150
+ "label": "DAEMON_HIGH",
151
+ "desc": "100 <= uid <= 999 (modern Linux service accounts)"
152
+ },
153
+ "4": {
154
+ "label": "OTHER",
155
+ "desc": "everything else (incl. uid >= 1001, nobody/65534)"
156
+ }
157
+ },
158
+ "f_dst_port_bucket": {
159
+ "0": {
160
+ "label": "NONE",
161
+ "desc": "row has no dst port (non-network event)"
162
+ },
163
+ "1": {
164
+ "label": "LOOPBACK",
165
+ "desc": "dport == 0 (Tetragon's loopback / unset sentinel)"
166
+ },
167
+ "2": {
168
+ "label": "SSH",
169
+ "desc": "dport == 22"
170
+ },
171
+ "3": {
172
+ "label": "DNS",
173
+ "desc": "dport == 53"
174
+ },
175
+ "4": {
176
+ "label": "WEB",
177
+ "desc": "dport in {80, 443, 8080, 8443}"
178
+ },
179
+ "5": {
180
+ "label": "PRIV",
181
+ "desc": "1 <= dport <= 1023 (reserved/privileged ports)"
182
+ },
183
+ "6": {
184
+ "label": "HIGH",
185
+ "desc": "dport >= 1024 (ephemeral / non-standard high ports)"
186
+ }
187
+ },
188
+ "f_args_length_bucket": {
189
+ "0": {
190
+ "label": "ZERO",
191
+ "desc": "no arguments"
192
+ },
193
+ "1": {
194
+ "label": "LT50",
195
+ "desc": "1 <= len(args) < 50 chars"
196
+ },
197
+ "2": {
198
+ "label": "RANGE_50_200",
199
+ "desc": "50 <= len(args) < 200 chars"
200
+ },
201
+ "3": {
202
+ "label": "RANGE_200_1K",
203
+ "desc": "200 <= len(args) < 1000 chars"
204
+ },
205
+ "4": {
206
+ "label": "RANGE_1K_5K",
207
+ "desc": "1000 <= len(args) < 5000 chars"
208
+ },
209
+ "5": {
210
+ "label": "GT5K",
211
+ "desc": "len(args) >= 5000 chars (suspect / oversize cmdlines)"
212
+ }
213
+ },
214
+ "f_cap_count_bucket": {
215
+ "0": {
216
+ "label": "ZERO",
217
+ "desc": "no Linux capabilities held"
218
+ },
219
+ "1": {
220
+ "label": "RANGE_1_2",
221
+ "desc": "1 or 2 capabilities"
222
+ },
223
+ "2": {
224
+ "label": "RANGE_3_5",
225
+ "desc": "3 to 5 capabilities"
226
+ },
227
+ "3": {
228
+ "label": "RANGE_6_15",
229
+ "desc": "6 to 15 capabilities"
230
+ },
231
+ "4": {
232
+ "label": "GE16",
233
+ "desc": "16 or more capabilities (essentially full root)"
234
+ }
235
+ },
236
+ "f_path_sens_cwd": {
237
+ "0": {
238
+ "label": "SENSITIVE_SYS",
239
+ "desc": "/etc/, /root/, /boot/, /usr/sbin/, /sbin/"
240
+ },
241
+ "1": {
242
+ "label": "STAGING",
243
+ "desc": "/tmp/, /var/tmp/, /dev/shm/"
244
+ },
245
+ "2": {
246
+ "label": "RUNTIME",
247
+ "desc": "/var/run/, /run/"
248
+ },
249
+ "3": {
250
+ "label": "USER_HOME",
251
+ "desc": "/home/, /Users/"
252
+ },
253
+ "4": {
254
+ "label": "PROC_SELF",
255
+ "desc": "/proc/self/ (own-process introspection)"
256
+ },
257
+ "5": {
258
+ "label": "VSCODE_DEV",
259
+ "desc": "developer-tooling paths: .vscode-server/, node_modules/, .git/"
260
+ },
261
+ "6": {
262
+ "label": "OTHER",
263
+ "desc": "anything else"
264
+ },
265
+ "7": {
266
+ "label": "NONE",
267
+ "desc": "row has no path (non-file event)"
268
+ }
269
+ },
270
+ "f_path_sens_binary": {
271
+ "0": {
272
+ "label": "SENSITIVE_SYS",
273
+ "desc": "/etc/, /root/, /boot/, /usr/sbin/, /sbin/"
274
+ },
275
+ "1": {
276
+ "label": "STAGING",
277
+ "desc": "/tmp/, /var/tmp/, /dev/shm/"
278
+ },
279
+ "2": {
280
+ "label": "RUNTIME",
281
+ "desc": "/var/run/, /run/"
282
+ },
283
+ "3": {
284
+ "label": "USER_HOME",
285
+ "desc": "/home/, /Users/"
286
+ },
287
+ "4": {
288
+ "label": "PROC_SELF",
289
+ "desc": "/proc/self/ (own-process introspection)"
290
+ },
291
+ "5": {
292
+ "label": "VSCODE_DEV",
293
+ "desc": "developer-tooling paths: .vscode-server/, node_modules/, .git/"
294
+ },
295
+ "6": {
296
+ "label": "OTHER",
297
+ "desc": "anything else"
298
+ },
299
+ "7": {
300
+ "label": "NONE",
301
+ "desc": "row has no path (non-file event)"
302
+ }
303
+ },
304
+ "f_path_sens_kp": {
305
+ "0": {
306
+ "label": "SENSITIVE_SYS",
307
+ "desc": "/etc/, /root/, /boot/, /usr/sbin/, /sbin/"
308
+ },
309
+ "1": {
310
+ "label": "STAGING",
311
+ "desc": "/tmp/, /var/tmp/, /dev/shm/"
312
+ },
313
+ "2": {
314
+ "label": "RUNTIME",
315
+ "desc": "/var/run/, /run/"
316
+ },
317
+ "3": {
318
+ "label": "USER_HOME",
319
+ "desc": "/home/, /Users/"
320
+ },
321
+ "4": {
322
+ "label": "PROC_SELF",
323
+ "desc": "/proc/self/ (own-process introspection)"
324
+ },
325
+ "5": {
326
+ "label": "VSCODE_DEV",
327
+ "desc": "developer-tooling paths: .vscode-server/, node_modules/, .git/"
328
+ },
329
+ "6": {
330
+ "label": "OTHER",
331
+ "desc": "anything else"
332
+ },
333
+ "7": {
334
+ "label": "NONE",
335
+ "desc": "row has no path (non-file event)"
336
+ }
337
+ },
338
+ "f_kp_fd_install_path_sensitivity": {
339
+ "0": {
340
+ "label": "SENSITIVE_SYS",
341
+ "desc": "/etc/, /root/, /boot/, /usr/sbin/, /sbin/"
342
+ },
343
+ "1": {
344
+ "label": "STAGING",
345
+ "desc": "/tmp/, /var/tmp/, /dev/shm/"
346
+ },
347
+ "2": {
348
+ "label": "RUNTIME",
349
+ "desc": "/var/run/, /run/"
350
+ },
351
+ "3": {
352
+ "label": "USER_HOME",
353
+ "desc": "/home/, /Users/"
354
+ },
355
+ "4": {
356
+ "label": "PROC_SELF",
357
+ "desc": "/proc/self/ (own-process introspection)"
358
+ },
359
+ "5": {
360
+ "label": "VSCODE_DEV",
361
+ "desc": "developer-tooling paths: .vscode-server/, node_modules/, .git/"
362
+ },
363
+ "6": {
364
+ "label": "OTHER",
365
+ "desc": "anything else"
366
+ },
367
+ "7": {
368
+ "label": "NONE",
369
+ "desc": "row has no path (non-file event)"
370
+ }
371
+ },
372
+ "f_kp_mmap_path_sensitivity": {
373
+ "0": {
374
+ "label": "SENSITIVE_SYS",
375
+ "desc": "/etc/, /root/, /boot/, /usr/sbin/, /sbin/"
376
+ },
377
+ "1": {
378
+ "label": "STAGING",
379
+ "desc": "/tmp/, /var/tmp/, /dev/shm/"
380
+ },
381
+ "2": {
382
+ "label": "RUNTIME",
383
+ "desc": "/var/run/, /run/"
384
+ },
385
+ "3": {
386
+ "label": "USER_HOME",
387
+ "desc": "/home/, /Users/"
388
+ },
389
+ "4": {
390
+ "label": "PROC_SELF",
391
+ "desc": "/proc/self/ (own-process introspection)"
392
+ },
393
+ "5": {
394
+ "label": "VSCODE_DEV",
395
+ "desc": "developer-tooling paths: .vscode-server/, node_modules/, .git/"
396
+ },
397
+ "6": {
398
+ "label": "OTHER",
399
+ "desc": "anything else"
400
+ },
401
+ "7": {
402
+ "label": "NONE",
403
+ "desc": "row has no path (non-file event)"
404
+ }
405
+ },
406
+ "f_cmdline_entropy": {
407
+ "0": {
408
+ "label": "NULL",
409
+ "desc": "row has no cmdline (non-exec event)"
410
+ },
411
+ "1": {
412
+ "label": "LOW",
413
+ "desc": "Shannon entropy < 2.0 (very repetitive / short)"
414
+ },
415
+ "2": {
416
+ "label": "MID_LOW",
417
+ "desc": "2.0 <= entropy < 3.5"
418
+ },
419
+ "3": {
420
+ "label": "MID",
421
+ "desc": "3.5 <= entropy < 4.5"
422
+ },
423
+ "4": {
424
+ "label": "MID_HIGH",
425
+ "desc": "4.5 <= entropy < 5.5"
426
+ },
427
+ "5": {
428
+ "label": "HIGH",
429
+ "desc": "entropy >= 5.5 (random-looking / encoded payloads)"
430
+ }
431
+ },
432
+ "f_cmdline_compress": {
433
+ "0": {
434
+ "label": "NULL",
435
+ "desc": "row has no cmdline"
436
+ },
437
+ "1": {
438
+ "label": "VERY_COMPRESSIBLE",
439
+ "desc": "ratio < 0.3"
440
+ },
441
+ "2": {
442
+ "label": "COMPRESSIBLE",
443
+ "desc": "0.3 <= ratio < 0.5"
444
+ },
445
+ "3": {
446
+ "label": "MODERATE",
447
+ "desc": "0.5 <= ratio < 0.7"
448
+ },
449
+ "4": {
450
+ "label": "WEAK",
451
+ "desc": "0.7 <= ratio < 0.9"
452
+ },
453
+ "5": {
454
+ "label": "INCOMPRESSIBLE",
455
+ "desc": "ratio >= 0.9 (random / pre-encrypted)"
456
+ }
457
+ },
458
+ "f_time_since_parent_exec": {
459
+ "0": {
460
+ "label": "NULL",
461
+ "desc": "no parent exec_time available"
462
+ },
463
+ "1": {
464
+ "label": "SUB_MS",
465
+ "desc": "< 1 ms since parent exec"
466
+ },
467
+ "2": {
468
+ "label": "MS",
469
+ "desc": "1 ms .. 1 s"
470
+ },
471
+ "3": {
472
+ "label": "SEC",
473
+ "desc": "1 s .. 1 min"
474
+ },
475
+ "4": {
476
+ "label": "MIN",
477
+ "desc": "1 min .. 1 h"
478
+ },
479
+ "5": {
480
+ "label": "HR",
481
+ "desc": ">= 1 h"
482
+ }
483
+ },
484
+ "f_kp_tcp_connect_sock_family": {
485
+ "0": {
486
+ "label": "AF_INET",
487
+ "desc": "IPv4"
488
+ },
489
+ "1": {
490
+ "label": "AF_INET6",
491
+ "desc": "IPv6"
492
+ },
493
+ "2": {
494
+ "label": "NONE",
495
+ "desc": "row is not a network event"
496
+ },
497
+ "3": {
498
+ "label": "OOV",
499
+ "desc": "Socket family outside {AF_INET, AF_INET6}"
500
+ }
501
+ },
502
+ "f_action_family": {
503
+ "0": {
504
+ "label": "PROC_EXEC",
505
+ "desc": "Process execution"
506
+ },
507
+ "1": {
508
+ "label": "PROC_EXIT",
509
+ "desc": "Process termination"
510
+ },
511
+ "2": {
512
+ "label": "NET_CONNECT",
513
+ "desc": "Outbound TCP connect()"
514
+ },
515
+ "3": {
516
+ "label": "NET_CLOSE",
517
+ "desc": "TCP socket close"
518
+ },
519
+ "4": {
520
+ "label": "NET_ACCEPT",
521
+ "desc": "Inbound TCP accept()"
522
+ },
523
+ "5": {
524
+ "label": "NET_SEND",
525
+ "desc": "UDP send (e.g. DNS at dport=53)"
526
+ },
527
+ "6": {
528
+ "label": "FILE_OPEN",
529
+ "desc": "File descriptor install (file open)"
530
+ },
531
+ "7": {
532
+ "label": "FILE_DELETE",
533
+ "desc": "File unlink"
534
+ },
535
+ "8": {
536
+ "label": "FILE_CHMOD",
537
+ "desc": "File permission change"
538
+ },
539
+ "9": {
540
+ "label": "PRIV_CHANGE",
541
+ "desc": "Credential change (setuid / setgid / capset)"
542
+ },
543
+ "10": {
544
+ "label": "PROC_INJECT_PTRACE",
545
+ "desc": "ptrace() — process injection / debugger"
546
+ },
547
+ "11": {
548
+ "label": "PROC_INJECT_VMWRITE",
549
+ "desc": "process_vm_writev() — cross-process memory write"
550
+ },
551
+ "12": {
552
+ "label": "MEM_MAP",
553
+ "desc": "mmap() of a file"
554
+ },
555
+ "13": {
556
+ "label": "MEM_PROTECT",
557
+ "desc": "mprotect() — memory permission change"
558
+ },
559
+ "14": {
560
+ "label": "OTHER",
561
+ "desc": "Catch-all (unknown kprobe in v0.2 corpus)"
562
+ }
563
+ },
564
+ "f_path_category": {
565
+ "0": {
566
+ "label": "VSCODE_DEV",
567
+ "desc": "developer-tooling paths (.vscode-server/, node_modules/, .git/)"
568
+ },
569
+ "1": {
570
+ "label": "SENSITIVE_SYS",
571
+ "desc": "/etc/, /root/, /boot/, /usr/sbin/, /sbin/"
572
+ },
573
+ "2": {
574
+ "label": "TEMP_STAGING",
575
+ "desc": "/tmp/, /var/tmp/, /dev/shm/"
576
+ },
577
+ "3": {
578
+ "label": "RUNTIME",
579
+ "desc": "/var/run/, /run/, /proc/"
580
+ },
581
+ "4": {
582
+ "label": "LIBRARY",
583
+ "desc": "/usr/lib/, /usr/lib32/, /usr/lib64/, /usr/local/lib/, /lib/, /lib32/, /lib64/ — NEW in 3a'"
584
+ },
585
+ "5": {
586
+ "label": "USR_BIN",
587
+ "desc": "/usr/bin/, /usr/local/bin/, /usr/local/sbin/ — NEW in 3a'"
588
+ },
589
+ "6": {
590
+ "label": "OPT",
591
+ "desc": "/opt/ — NEW in 3a'"
592
+ },
593
+ "7": {
594
+ "label": "USER_HOME",
595
+ "desc": "/home/, /Users/"
596
+ },
597
+ "8": {
598
+ "label": "PROC_SELF",
599
+ "desc": "/proc/self/"
600
+ },
601
+ "9": {
602
+ "label": "OTHER",
603
+ "desc": "anything else"
604
+ },
605
+ "10": {
606
+ "label": "NONE",
607
+ "desc": "row has no path"
608
+ }
609
+ },
610
+ "f_dst_ip_category": {
611
+ "0": {
612
+ "label": "LOOPBACK",
613
+ "desc": "127.0.0.0/8 or ::1"
614
+ },
615
+ "1": {
616
+ "label": "LINK_LOCAL",
617
+ "desc": "169.254.0.0/16 or fe80::/10"
618
+ },
619
+ "2": {
620
+ "label": "PRIVATE",
621
+ "desc": "RFC1918: 10/8, 172.16/12, 192.168/16"
622
+ },
623
+ "3": {
624
+ "label": "MULTICAST",
625
+ "desc": "224.0.0.0/4 or ff00::/8"
626
+ },
627
+ "4": {
628
+ "label": "PUBLIC",
629
+ "desc": "publicly-routable IP (anonymized at parser; only category survives)"
630
+ },
631
+ "5": {
632
+ "label": "NONE",
633
+ "desc": "row has no dst IP"
634
+ }
635
+ },
636
+ "f_dst_port_category": {
637
+ "0": {
638
+ "label": "DNS",
639
+ "desc": "dport == 53"
640
+ },
641
+ "1": {
642
+ "label": "WEB",
643
+ "desc": "dport in {80, 443, 8080, 8443}"
644
+ },
645
+ "2": {
646
+ "label": "SSH",
647
+ "desc": "dport == 22"
648
+ },
649
+ "3": {
650
+ "label": "WELL_KNOWN",
651
+ "desc": "1 <= dport < 1024 (other reserved/privileged ports)"
652
+ },
653
+ "4": {
654
+ "label": "REGISTERED",
655
+ "desc": "1024 <= dport < 49152"
656
+ },
657
+ "5": {
658
+ "label": "HIGH",
659
+ "desc": "dport >= 49152 (ephemeral)"
660
+ },
661
+ "6": {
662
+ "label": "NONE",
663
+ "desc": "row has no dst port (non-network event)"
664
+ }
665
+ },
666
+ "f_object_category": {
667
+ "0": {
668
+ "label": "PROCESS",
669
+ "desc": "process_exec / process_exit"
670
+ },
671
+ "1": {
672
+ "label": "FILE",
673
+ "desc": "file-system kprobe (open/delete/chmod)"
674
+ },
675
+ "2": {
676
+ "label": "NETWORK",
677
+ "desc": "TCP/UDP socket kprobe"
678
+ },
679
+ "3": {
680
+ "label": "MEMORY",
681
+ "desc": "mmap / mprotect"
682
+ },
683
+ "4": {
684
+ "label": "CREDENTIAL",
685
+ "desc": "commit_creds (setuid/setgid/capset)"
686
+ },
687
+ "5": {
688
+ "label": "INJECT",
689
+ "desc": "ptrace / process_vm_writev"
690
+ },
691
+ "6": {
692
+ "label": "OTHER",
693
+ "desc": "kprobe outside the v0.2 vocabulary"
694
+ },
695
+ "7": {
696
+ "label": "NONE",
697
+ "desc": "row has no event_type"
698
+ }
699
+ }
700
+ },
701
+ "categorical_string_columns": {
702
+ "path_category": {
703
+ "VSCODE_DEV": "developer-tooling paths (.vscode-server/, node_modules/, .git/)",
704
+ "SENSITIVE_SYS": "/etc/, /root/, /boot/, /usr/sbin/, /sbin/",
705
+ "TEMP_STAGING": "/tmp/, /var/tmp/, /dev/shm/",
706
+ "RUNTIME": "/var/run/, /run/, /proc/",
707
+ "LIBRARY": "/usr/lib/, /usr/lib32/, /usr/lib64/, /usr/local/lib/, /lib/, /lib32/, /lib64/ (new in 3a')",
708
+ "USR_BIN": "/usr/bin/, /usr/local/bin/, /usr/local/sbin/ (new in 3a')",
709
+ "OPT": "/opt/ (new in 3a')",
710
+ "USER_HOME": "/home/, /Users/",
711
+ "PROC_SELF": "/proc/self/",
712
+ "OTHER": "anything else",
713
+ "NONE": "row has no path"
714
+ },
715
+ "object_category": {
716
+ "PROCESS": "process_exec / process_exit",
717
+ "FILE": "file-system kprobe (open/delete/chmod)",
718
+ "NETWORK": "TCP/UDP socket kprobe",
719
+ "CREDENTIAL": "commit_creds (setuid/setgid/capset)",
720
+ "INJECT": "ptrace / process_vm_writev",
721
+ "MEMORY": "mmap / mprotect",
722
+ "OTHER": "kprobe outside the v0.2 vocabulary",
723
+ "NONE": "row has no event_type"
724
+ },
725
+ "dst_ip_category": {
726
+ "LOOPBACK": "127.0.0.0/8 or ::1",
727
+ "LINK_LOCAL": "169.254.0.0/16 or fe80::/10",
728
+ "PRIVATE": "RFC1918: 10/8, 172.16/12, 192.168/16",
729
+ "MULTICAST": "224.0.0.0/4 or ff00::/8",
730
+ "PUBLIC": "publicly-routable IP (anonymized at parser; only category survives)",
731
+ "NONE": "row has no dst IP"
732
+ },
733
+ "dst_port_category": {
734
+ "DNS": "dport == 53",
735
+ "WEB": "dport in {80, 443, 8080, 8443}",
736
+ "SSH": "dport == 22",
737
+ "WELL_KNOWN": "1 <= dport < 1024 (other reserved/privileged ports)",
738
+ "REGISTERED": "1024 <= dport < 49152",
739
+ "HIGH": "dport >= 49152 (ephemeral)",
740
+ "NONE": "row has no dst port (non-network event)"
741
+ }
742
+ },
743
+ "boolean_flag_columns": {
744
+ "f_is_procfs_walk": "1 if the parser flagged this process_exec as a procfs-walk synthetic event; tag-not-drop policy means these flow through with the flag preserved (1,250 train + 96 test events have value=1).",
745
+ "f_uid_eq_parent": "1 if the process and its parent share the same uid.",
746
+ "f_is_setuid_exec": "1 if the exec was a setuid (uid != parent uid AND was a process_exec).",
747
+ "f_kp_commit_creds_uid_change": "1 on commit_creds events where the uid changed.",
748
+ "f_kp_commit_creds_cap_change": "1 on commit_creds events where the capability set changed.",
749
+ "f_kp_udp_sendmsg_dport_eq_53": "1 on udp_sendmsg events with dport=53 (DNS)."
750
+ },
751
+ "numeric_columns": {
752
+ "f_kp_fd_install_fd_int32": "Raw fd integer from fd_install (int32).",
753
+ "f_kp_mmap_prot_uint": "mmap prot bitmask (uint32).",
754
+ "f_kp_mprotect_prot_uint": "mprotect prot bitmask (uint64).",
755
+ "delta_t_prev_ns": "Nanoseconds since the previous event for the same proc_exec_id (per-process keying landed in the 2026-04-29 rebuild; previously global, which produced 44,947 train + 1,643 test negative residuals — now zero).",
756
+ "process_age_ns": "Nanoseconds since the process started (event_time - proc_start_time)."
757
+ },
758
+ "bucketed_int_columns": {
759
+ "f_delta_t_log_bucket": "Half-decade bucket of log10(delta_t_prev_ns + 1).",
760
+ "f_process_age_log_bucket": "Half-decade bucket of log10(process_age_ns + 1).",
761
+ "f_lineage_depth": "Walked-ancestors depth at parser time (capped per Plan G)."
762
+ },
763
+ "hash_columns": [
764
+ {
765
+ "column": "f_proc_name_hash",
766
+ "buckets": 2048,
767
+ "source_in_raw": "proc_binary (basename)",
768
+ "desc": "Hash of the executable basename (e.g. 'bash', 'curl'). Bumped 1024 -> 2048 in 3a' (audit cell 43 showed 27.7% collision on the old size)."
769
+ },
770
+ {
771
+ "column": "f_parent_proc_hash",
772
+ "buckets": 1024,
773
+ "source_in_raw": "parent_binary (basename)",
774
+ "desc": "Hash of the parent process basename."
775
+ },
776
+ {
777
+ "column": "f_proc_cwd_hash",
778
+ "buckets": 4096,
779
+ "source_in_raw": "proc_cwd (sanitized at release; see anonymization log)",
780
+ "desc": "Hash of the first FOUR path components of proc_cwd. Capacity bumped 256 -> 4096 AND depth 2 -> 4 in 3a' (audit cell 43 showed 90.2% collision on the old depth-2 / 256-bucket sizing). Depth-4 captures per-project granularity."
781
+ },
782
+ {
783
+ "column": "f_lineage_bag_hash",
784
+ "buckets": 256,
785
+ "source_in_raw": "(composite — sorted set of lineage basenames)",
786
+ "desc": "Hash of the bag-of-ancestors basename set; order-invariant."
787
+ },
788
+ {
789
+ "column": "f_parent_child_pair_hash",
790
+ "buckets": 1024,
791
+ "source_in_raw": "(derived: parent_binary_basename + ' -> ' + proc_binary_basename)",
792
+ "desc": "Hash of the (parent_basename, child_basename) pair string."
793
+ },
794
+ {
795
+ "column": "f_root_ancestor_basename_hash",
796
+ "buckets": 1024,
797
+ "source_in_raw": "(derived: lineage walk to the process tree root)",
798
+ "desc": "Hash of the root-of-tree basename (typically systemd / sshd)."
799
+ },
800
+ {
801
+ "column": "f_process_tree_id_hash",
802
+ "buckets": 4096,
803
+ "source_in_raw": "(derived: stable hash of the root proc_exec_id)",
804
+ "desc": "Identifier for the process subtree this event belongs to."
805
+ }
806
+ ],
807
+ "identifier_columns": {
808
+ "event_time": "RFC3339 nanosecond timestamp (timestamp[ns, tz=UTC]).",
809
+ "proc_exec_id": "Salted-HMAC pseudonym of the original Tetragon exec_id (format 'exec_<32hex>'). Equality joins are preserved — lineage walks still work — but the original 'host:ts:pid' payload is erased.",
810
+ "proc_pid": "OS process id at the time of the event (uint32). Container-internal."
811
+ },
812
+ "side_string_columns": {
813
+ "proc_binary_basename": "Basename of the executable. Aliased per the v0.2.1 anonymizer: allowlisted distro / dev-tool basenames pass through verbatim; everything else becomes 'custom_tool_NNN' (lex-sorted) or the fixed 'vscode_server_bin' alias for VSCode commit-hash binaries. See reproduction/alias_map.json.",
814
+ "parent_binary_basename": "Basename of the parent executable. Same alias map as proc_binary_basename.",
815
+ "proc_cwd_normalized": "Sanitized current working directory: '/home/<user>' -> '/home/_user_', '/mnt/projects/<X>' -> '/workspace/_project_', etc. See anonymization log for the full rule set.",
816
+ "root_ancestor_basename": "Basename of the process-tree root. Same alias map as the other basename columns.",
817
+ "process_tree_root_exec_id": "Salted-HMAC pseudonym of the original Tetragon exec_id of the process-tree root (format 'exec_<32hex>').",
818
+ "parent_child_pair": "Parent->child basename pair, re-composed from the aliased basenames after Step B of the anonymizer (component-first guarantee — aliased value flows through).",
819
+ "token": "Composite event identifier of shape '<action_family>:<basename>[:<category>...]'. Re-generated via the canonical format_token() from the aliased basenames + already-bucketed categories, so the original (unaliased) basename never re-enters the token."
820
+ }
821
+ }
reproduction/requirements.txt ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Minimal requirements for parser and baseline infrastructure
2
+ # Add others as we implement more features
3
+
4
+ # Testing
5
+ pytest>=7.0
6
+
7
+ # Parser (v0.2 tetragon-native)
8
+ pyarrow>=15.0
9
+ PyYAML>=6.0
10
+
11
+ # Data processing (future)
12
+ # numpy
13
+ # pandas
14
+
15
+ # Model inference (future)
16
+ # torch
17
+ # mamba-ssm
18
+
19
+ # Plotting/visualization (notebooks)
20
+ # matplotlib
21
+ # jupyter
22
+
23
+ # Code quality
24
+ ruff>=0.1.0
reproduction/tetragon-configs/identity_map.yaml ADDED
@@ -0,0 +1,81 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Identity map for Tetragon telemetry on <YOUR_HOST> / LXC container 110.
2
+ #
3
+ # IMPORTANT — container-internal UIDs, not host-offset UIDs.
4
+ # Empirically, Tetragon 1.6.1 on <YOUR_HOST> reports the container-internal UID on
5
+ # events emitted from inside LXC container 110 (e.g. <YOUR_USER> shows up as uid
6
+ # 1000, not 101000). The earlier subuid-offset hypothesis (host_uid =
7
+ # 100000 + container_uid) was incorrect for this Tetragon build; verified
8
+ # against 156k events in data/raw/events-2026-04-14T03-53-05.161.json.gz.
9
+ #
10
+ # Enumeration command (re-run if container users change):
11
+ # pct exec 110 -- cat /etc/passwd \
12
+ # | awk -F: '{printf " %d: %s # %s\n", $3, $1, $1}'
13
+ #
14
+ # Parser lookup is keyed on the UID Tetragon emits in process.uid.
15
+ # Unknown UIDs (not in this map) fall through to persona `unknown` with
16
+ # identity_relation_to_host = unknown_user.
17
+
18
+ personas:
19
+ # ---- Primary actors (high behavioral signal) ----
20
+ 0: container_root # root — privileged actor, runs system changes
21
+ 1000: primary_user # <YOUR_USER> — the human primary user
22
+
23
+ # ---- Package manager (critical for demo attack chain) ----
24
+ 42: service_apt # _apt
25
+
26
+ # ---- Active service accounts (distinct behavioral baselines) ----
27
+ 33: service_www # www-data
28
+ 103: service_ssh # sshd
29
+ 104: service_postfix # postfix
30
+ 101: service_dbus # messagebus
31
+ 102: service_syslog # syslog
32
+ 100: service_dhcp # dhcpcd
33
+ 106: service_tcpdump # tcpdump
34
+ 990: service_polkit # polkitd
35
+
36
+ # ---- systemd sub-daemons (collapse under one label; similar behavior) ----
37
+ 998: service_systemd # systemd-network
38
+ 997: service_systemd # systemd-timesync
39
+ 992: service_systemd # systemd-resolve
40
+
41
+ # ---- Historical Unix system accounts (rarely execute; collapse) ----
42
+ # Kept explicit so unexpected activity still attributes correctly rather
43
+ # than falling into `unknown`.
44
+ 1: service_system # daemon
45
+ 2: service_system # bin
46
+ 3: service_system # sys
47
+ 4: service_system # sync
48
+ 5: service_system # games
49
+ 6: service_system # man
50
+ 7: service_system # lp
51
+ 8: service_system # mail
52
+ 9: service_system # news
53
+ 10: service_system # uucp
54
+ 13: service_system # proxy
55
+ 34: service_system # backup
56
+ 38: service_system # list
57
+ 39: service_system # irc
58
+ 105: service_system # uuidd
59
+
60
+ # ---- Nobody / overflow UID ----
61
+ 65534: service_nobody
62
+
63
+ # Persona -> schema field derivations (consumed by the parser).
64
+ # Maps each persona label to the v1.3 event-contract enum values for
65
+ # user_role and identity_relation_to_host.
66
+ persona_roles:
67
+ container_root: { user_role: admin, relation: admin_support }
68
+ primary_user: { user_role: admin, relation: primary_user }
69
+ service_apt: { user_role: service, relation: service_account }
70
+ service_www: { user_role: service, relation: service_account }
71
+ service_ssh: { user_role: service, relation: service_account }
72
+ service_postfix: { user_role: service, relation: service_account }
73
+ service_dbus: { user_role: service, relation: service_account }
74
+ service_syslog: { user_role: service, relation: service_account }
75
+ service_dhcp: { user_role: service, relation: service_account }
76
+ service_tcpdump: { user_role: service, relation: service_account }
77
+ service_polkit: { user_role: service, relation: service_account }
78
+ service_systemd: { user_role: system, relation: service_account }
79
+ service_system: { user_role: system, relation: service_account }
80
+ service_nobody: { user_role: system, relation: service_account }
81
+ unknown: { user_role: unknown, relation: unknown_user }
reproduction/tetragon-configs/tetragon/attribute-monitor.yaml ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # File attribute monitoring TracingPolicy for Mamba behavioral telemetry
2
+ # Captures chmod / permission changes from LXC containers (excludes host events)
3
+ # Deploy to <YOUR_HOST>: cp attribute-monitor.yaml /etc/tetragon/tetragon.tp.d/
4
+ #
5
+ # Key use cases:
6
+ # - Payload preparation (T1222 File and Directory Permissions Modification)
7
+ # - Classic attack chain: curl payload -> chmod +x -> execute
8
+ #
9
+ # Behavioral signal rationale:
10
+ # chmod_common is the single kernel function all permission-change syscalls
11
+ # (chmod, fchmod, fchmodat) flow through. Capturing it gives the Mamba model
12
+ # a "preparation" state that precedes execution. Sequences like
13
+ # (network, connect) -> (file, write) -> (file, modify[chmod +x]) -> (system, process, spawn)
14
+ # are a high-signal attack-chain pattern that is much easier for an SSM to
15
+ # flag than the exec event in isolation.
16
+
17
+ apiVersion: cilium.io/v1alpha1
18
+ kind: TracingPolicy
19
+ metadata:
20
+ name: "attribute-monitor"
21
+ spec:
22
+ kprobes:
23
+ # File permission / mode change (T1222)
24
+ # Parser maps this to (file, file, modify) with mode bits in evidence.
25
+ # Evidence field captures the mode so downstream logic can identify
26
+ # grants of the execute bit (0o111) specifically.
27
+ # NOTE: chmod_common's mode arg is kernel type umode_t (16-bit unsigned).
28
+ # uint16 silences the type-mismatch warning Tetragon emits on load.
29
+ - call: "chmod_common"
30
+ syscall: false
31
+ args:
32
+ - index: 0
33
+ type: "path" # target file path
34
+ - index: 1
35
+ type: "uint16" # new mode bits (umode_t)
36
+ selectors:
37
+ - matchNamespaces:
38
+ - namespace: "Pid"
39
+ operator: "NotIn"
40
+ values:
41
+ - "host_ns"
reproduction/tetragon-configs/tetragon/file-monitor.yaml ADDED
@@ -0,0 +1,46 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # File monitoring TracingPolicy for Mamba behavioral telemetry
2
+ # Captures file opens and deletes from LXC containers (excludes host events)
3
+ # Deploy to <YOUR_HOST>: cp file-monitor.yaml /etc/tetragon/tetragon.tp.d/
4
+ #
5
+ # Key use cases:
6
+ # - ART marker detection (/tmp/art_start_*, /tmp/art_end_*)
7
+ # - File access patterns for behavioral modeling
8
+ # - Sensitive file access (e.g., /etc/shadow, /etc/passwd)
9
+
10
+ apiVersion: cilium.io/v1alpha1
11
+ kind: TracingPolicy
12
+ metadata:
13
+ name: "file-monitor"
14
+ spec:
15
+ kprobes:
16
+ # File open — fires when a file descriptor is installed (successful open)
17
+ # Tetragon extracts the full file path from the kernel file struct
18
+ - call: "fd_install"
19
+ syscall: false
20
+ args:
21
+ - index: 0
22
+ type: "int" # file descriptor number
23
+ - index: 1
24
+ type: "file" # file struct — Tetragon extracts path
25
+ selectors:
26
+ - matchNamespaces:
27
+ - namespace: "Pid"
28
+ operator: "NotIn"
29
+ values:
30
+ - "host_ns"
31
+
32
+ # File delete — fires on unlink/unlinkat syscall
33
+ # Captures file path being deleted
34
+ - call: "do_unlinkat"
35
+ syscall: false
36
+ args:
37
+ - index: 0
38
+ type: "int" # directory fd
39
+ - index: 1
40
+ type: "string" # pathname
41
+ selectors:
42
+ - matchNamespaces:
43
+ - namespace: "Pid"
44
+ operator: "NotIn"
45
+ values:
46
+ - "host_ns"
reproduction/tetragon-configs/tetragon/injection-monitor.yaml ADDED
@@ -0,0 +1,82 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Process injection monitoring TracingPolicy for Mamba behavioral telemetry
2
+ # Captures ptrace, mprotect (RX/RWX flips), and process_vm_writev from LXC containers
3
+ # Deploy to <YOUR_HOST>: cp injection-monitor.yaml /etc/tetragon/tetragon.tp.d/
4
+ #
5
+ # Key use cases:
6
+ # - Process injection via ptrace (T1055.008)
7
+ # - Shellcode activation via RW->RX mprotect flip (T1055)
8
+ # - Cross-process memory write injection (T1055.009)
9
+ #
10
+ # Behavioral signal rationale:
11
+ # memory-monitor.yaml captures mmap(PROT_EXEC) — naive in-memory execution.
12
+ # Sophisticated attackers evade that by allocating RW, writing payload, then
13
+ # flipping the page to RX via mprotect. security_file_mprotect with a
14
+ # PROT_EXEC mask catches that evasion path. Filtering on PROT_EXEC keeps
15
+ # the event volume sane — routine mprotect calls (heap management, PROT_NONE
16
+ # guard pages) are ignored.
17
+
18
+ apiVersion: cilium.io/v1alpha1
19
+ kind: TracingPolicy
20
+ metadata:
21
+ name: "injection-monitor"
22
+ spec:
23
+ kprobes:
24
+ # Process debugging and ptrace-based injection (T1055.008)
25
+ # request arg identifies PTRACE_ATTACH, PTRACE_POKEDATA, PTRACE_DETACH, etc.
26
+ # Parser maps this to action=trace (v1.3 schema).
27
+ # NOTE: Tetragon's kprobe type vocabulary does not include "long" — use int64.
28
+ - call: "sys_ptrace"
29
+ syscall: true
30
+ args:
31
+ - index: 0
32
+ type: "int64" # ptrace request (PTRACE_ATTACH=16, POKEDATA=5, etc.)
33
+ - index: 1
34
+ type: "int" # target PID
35
+ selectors:
36
+ - matchNamespaces:
37
+ - namespace: "Pid"
38
+ operator: "NotIn"
39
+ values:
40
+ - "host_ns"
41
+
42
+ # Memory protection flip (T1055)
43
+ # Only fires when PROT_EXEC is being requested — captures the RW->RX
44
+ # shellcode activation pattern. Parser maps this to action=protect (v1.3).
45
+ #
46
+ # NOTE: security_file_mprotect's first arg is a struct vm_area_struct*
47
+ # which Tetragon's kprobe type vocabulary does not expose directly.
48
+ # We skip arg[0] and only extract the protection flags — Tetragon still
49
+ # captures the calling process's identity from event metadata.
50
+ - call: "security_file_mprotect"
51
+ syscall: false
52
+ args:
53
+ - index: 1
54
+ type: "uint64" # reqprot (requested protection flags)
55
+ - index: 2
56
+ type: "uint64" # prot (actual protection applied)
57
+ selectors:
58
+ - matchNamespaces:
59
+ - namespace: "Pid"
60
+ operator: "NotIn"
61
+ values:
62
+ - "host_ns"
63
+ matchArgs:
64
+ - index: 1
65
+ operator: "Mask"
66
+ values:
67
+ - "4" # PROT_EXEC — only capture flips granting execute
68
+
69
+ # Cross-process memory write (T1055.009)
70
+ # Faster injection path that does not require PTRACE_ATTACH.
71
+ # Parser maps this to action=write with object_type=process.
72
+ - call: "sys_process_vm_writev"
73
+ syscall: true
74
+ args:
75
+ - index: 0
76
+ type: "int" # target PID
77
+ selectors:
78
+ - matchNamespaces:
79
+ - namespace: "Pid"
80
+ operator: "NotIn"
81
+ values:
82
+ - "host_ns"
reproduction/tetragon-configs/tetragon/logrotate.tetragon ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Logrotate config for Tetragon JSON event stream on <YOUR_HOST>.
2
+ # Install to: /etc/logrotate.d/tetragon
3
+ #
4
+ # Rotation strategy:
5
+ # - Rotate daily OR when file exceeds 500MB (whichever first)
6
+ # - Keep 14 compressed rotations (~2 weeks of history)
7
+ # - copytruncate avoids Tetragon needing to reopen the file
8
+ # (Tetragon writes via append — truncate-in-place is safe)
9
+ #
10
+ # Install:
11
+ # sudo cp mamba-edge/config/tetragon/logrotate.tetragon /etc/logrotate.d/tetragon
12
+ # sudo logrotate -d /etc/logrotate.d/tetragon # dry-run to verify
13
+ # sudo logrotate -f /etc/logrotate.d/tetragon # force initial rotation
14
+
15
+ /var/log/tetragon/events.json {
16
+ daily
17
+ maxsize 500M
18
+ rotate 14
19
+ compress
20
+ delaycompress
21
+ missingok
22
+ notifempty
23
+ copytruncate
24
+ create 0640 root root
25
+ }
reproduction/tetragon-configs/tetragon/memory-monitor.yaml ADDED
@@ -0,0 +1,39 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Memory monitoring TracingPolicy for Mamba behavioral telemetry
2
+ # Detects fileless execution and in-memory attack techniques from LXC containers
3
+ # Deploy to <YOUR_HOST>: cp memory-monitor.yaml /etc/tetragon/tetragon.tp.d/
4
+ #
5
+ # Key use cases:
6
+ # - Fileless malware (shellcode injection into anonymous executable memory)
7
+ # - Process hollowing (unmap legitimate code, map malicious code)
8
+ # - Living-off-the-land (reflective DLL/SO loading)
9
+ # - MITRE ATT&CK: T1055 (Process Injection), T1620 (Reflective Code Loading)
10
+
11
+ apiVersion: cilium.io/v1alpha1
12
+ kind: TracingPolicy
13
+ metadata:
14
+ name: "memory-monitor"
15
+ spec:
16
+ kprobes:
17
+ # Memory-mapped file operations — LSM hook that fires on mmap calls
18
+ # Captures both file-backed mappings (normal .so loads) and anonymous
19
+ # executable mappings (fileless execution signal).
20
+ # The file arg is null for anonymous mappings — the parser uses this
21
+ # to distinguish normal library loads from suspicious in-memory execution.
22
+ - call: "security_mmap_file"
23
+ syscall: false
24
+ args:
25
+ - index: 0
26
+ type: "file" # file struct (null for anonymous mappings)
27
+ - index: 1
28
+ type: "uint32" # protection flags (PROT_READ=1, PROT_WRITE=2, PROT_EXEC=4)
29
+ selectors:
30
+ - matchNamespaces:
31
+ - namespace: "Pid"
32
+ operator: "NotIn"
33
+ values:
34
+ - "host_ns"
35
+ matchArgs:
36
+ - index: 1
37
+ operator: "Mask"
38
+ values:
39
+ - "4" # PROT_EXEC — only capture mappings with execute permission
reproduction/tetragon-configs/tetragon/network-monitor.yaml ADDED
@@ -0,0 +1,80 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Network monitoring TracingPolicy for Mamba behavioral telemetry
2
+ # Captures TCP connections and DNS queries from LXC containers (excludes host events)
3
+ # Deploy to <YOUR_HOST>: cp network-monitor.yaml /etc/tetragon/tetragon.tp.d/
4
+ #
5
+ # Key use cases:
6
+ # - C2 callback detection (unusual outbound connections)
7
+ # - Lateral movement detection
8
+ # - DNS resolution monitoring (normal domains vs C2/DGA domains)
9
+ # - Normal network behavior baseline
10
+
11
+ apiVersion: cilium.io/v1alpha1
12
+ kind: TracingPolicy
13
+ metadata:
14
+ name: "network-monitor"
15
+ spec:
16
+ kprobes:
17
+ # Outbound TCP connection — fires when a TCP connection is initiated
18
+ # Tetragon extracts src/dst IP:port from the sock struct
19
+ - call: "tcp_connect"
20
+ syscall: false
21
+ args:
22
+ - index: 0
23
+ type: "sock" # socket struct — Tetragon extracts family, src/dst IP:port
24
+ selectors:
25
+ - matchNamespaces:
26
+ - namespace: "Pid"
27
+ operator: "NotIn"
28
+ values:
29
+ - "host_ns"
30
+
31
+ # TCP connection close — captures connection lifecycle end
32
+ # Useful for connection duration analysis
33
+ - call: "tcp_close"
34
+ syscall: false
35
+ args:
36
+ - index: 0
37
+ type: "sock" # socket struct
38
+ selectors:
39
+ - matchNamespaces:
40
+ - namespace: "Pid"
41
+ operator: "NotIn"
42
+ values:
43
+ - "host_ns"
44
+
45
+ # Inbound TCP accept — fires when a listening socket accepts a connection
46
+ # Detects services receiving connections (e.g., reverse shells, bind shells)
47
+ - call: "inet_csk_accept"
48
+ syscall: false
49
+ return: true
50
+ returnArg:
51
+ index: 0
52
+ type: "sock" # accepted socket with peer info
53
+ selectors:
54
+ - matchNamespaces:
55
+ - namespace: "Pid"
56
+ operator: "NotIn"
57
+ values:
58
+ - "host_ns"
59
+
60
+ # DNS queries — captures UDP sends to port 53
61
+ # High-signal for C2 detection: model learns normal resolution patterns
62
+ # (github.com, apt repos) vs anomalous (DGA domains, unusual resolvers)
63
+ - call: "udp_sendmsg"
64
+ syscall: false
65
+ args:
66
+ - index: 0
67
+ type: "sock" # socket struct — dst IP:port
68
+ - index: 1
69
+ type: "int" # message length
70
+ selectors:
71
+ - matchNamespaces:
72
+ - namespace: "Pid"
73
+ operator: "NotIn"
74
+ values:
75
+ - "host_ns"
76
+ matchArgs:
77
+ - index: 0
78
+ operator: "DPort"
79
+ values:
80
+ - "53"
reproduction/tetragon-configs/tetragon/privilege-monitor.yaml ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Privilege escalation TracingPolicy for Mamba behavioral telemetry
2
+ # Captures credential changes from LXC containers (excludes host events)
3
+ # Deploy to <YOUR_HOST>: cp privilege-monitor.yaml /etc/tetragon/tetragon.tp.d/
4
+ #
5
+ # Key use cases:
6
+ # - setuid/setgid detection
7
+ # - Capability changes (capset)
8
+ # - Privilege escalation attack detection (T1548)
9
+
10
+ apiVersion: cilium.io/v1alpha1
11
+ kind: TracingPolicy
12
+ metadata:
13
+ name: "privilege-monitor"
14
+ spec:
15
+ kprobes:
16
+ # Credential commit — ALL credential changes flow through this function
17
+ # Covers: setuid, setgid, setreuid, setresuid, capset, execve with suid bit
18
+ # Tetragon extracts uid, gid, and capabilities from the cred struct
19
+ - call: "commit_creds"
20
+ syscall: false
21
+ args:
22
+ - index: 0
23
+ type: "cred" # credential struct — uid, gid, caps
24
+ selectors:
25
+ - matchNamespaces:
26
+ - namespace: "Pid"
27
+ operator: "NotIn"
28
+ values:
29
+ - "host_ns"
reproduction/tetragon-configs/tetragon_parser.yaml ADDED
@@ -0,0 +1,43 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Runtime configuration for the Tetragon parser.
2
+ #
3
+ # Loaded by TetragonParserConfig.from_yaml() at parser construction.
4
+ # Keep secrets out of this file — user_id_salt is loaded from env in production.
5
+
6
+ host_id: <YOUR_HOST>
7
+ container_ctid: 110 # LXC container holding the monitored workload
8
+
9
+ # Salt for derive_user_id_hash(). Dev default is committed for reproducibility;
10
+ # production MUST override via TETRAGON_PARSER_SALT environment variable.
11
+ user_id_salt: dev_only_not_a_secret
12
+
13
+ identity_map_path: config/identity_map.yaml
14
+
15
+ # Container isolation filter.
16
+ #
17
+ # Tetragon 1.6.1 does not emit string-form cgroup / docker / pod fields on
18
+ # container events (all null in practice), so substring matching on cgroup
19
+ # paths is not a viable isolation mechanism. The PID-namespace inum IS
20
+ # present on every container event and uniquely identifies an LXC container
21
+ # within a single boot, so we pin on that instead.
22
+ #
23
+ # Value below is the observed PID-ns inum for container 110 as of 2026-04-14
24
+ # (verified against data/raw/events-2026-04-14T03-53-05.161.json.gz: 61k
25
+ # events attributable to this inum; sibling inums 4026533679 and 4026533788
26
+ # are rejected).
27
+ #
28
+ # BRITTLENESS: the inum changes every time container 110 is restarted.
29
+ # If the container is restarted, re-capture with:
30
+ # zcat <latest-event-archive> | python3 -c "import json,sys; [print(d['process_exec']['process']['ns']['pid'].get('inum')) for l in sys.stdin for d in [json.loads(l)] if 'process_exec' in d][:1]"
31
+ # then update this field.
32
+ #
33
+ # Set to null to accept any non-host PID namespace (dev/testing only).
34
+ target_pid_ns_inum: <YOUR_PIDNS_INUM>
35
+
36
+ # ART marker prefix. The parser matches process.arguments or kprobe file path
37
+ # against "<prefix>(start|end)_<technique>_<trial>" and emits a window label.
38
+ art_marker_prefix: "/tmp/art_"
39
+
40
+ # Timezone for derive_ts_bucket(). Bucket boundaries (workhours vs afterhours
41
+ # vs weekend) are computed in local time so behavioral baselines align with
42
+ # human work rhythms rather than UTC.
43
+ timezone: America/New_York