---
license: other
license_name: csil-3.0
license_link: https://catqualia.com/licensing
task_categories:
  - text-generation
language:
  - en
tags:
  - security
  - offensive-security
  - mitre-attack
  - mitigation
  - adversarial
  - instruction-tuning
pretty_name: Attack-Mitigation-Corpus — described attack techniques paired with mitigations
size_categories:
  - n<1K
---

# Attack-Mitigation-Corpus v1

One JSONL file of 123 security records. Each record is either a described **attack
technique** or a **mitigation**, and each attack record is paired with a mitigation record
covering a related defensive concern.

## Measured composition

| metric | value | command |
|---|---|---|
| rows | 123 | `wc -l < attack_mitigation_corpus.jsonl` |
| `kind == "attack"` | 106 | `python3 -c "import json,collections;print(collections.Counter(json.loads(l)['kind'] for l in open('attack_mitigation_corpus.jsonl')))"` |
| `kind == "mitigation"` | 17 | as above |
| attack rows with a `mitre_id` | 106 | `python3 -c "import json;print(sum(1 for l in open('attack_mitigation_corpus.jsonl') if json.loads(l).get('mitre_id')))"` |
| attack rows carrying a `paired_mitigation` object | 106 | `python3 -c "import json;print(sum(1 for l in open('attack_mitigation_corpus.jsonl') if 'paired_mitigation' in json.loads(l)))"` |
| attack rows with a short letter `grade` (A / B+ / B) | 71 | see below |
| attack rows whose `grade` is a long descriptive string instead | 35 | see below |
| distinct `technique` values | 106 | `python3 -c "import json;print(len({json.loads(l).get('technique') for l in open('attack_mitigation_corpus.jsonl') if json.loads(l).get('kind')=='attack'}))"` |
| distinct `phase` values | 34 | as above on `phase` |

## Schema

| field | type | rows containing it |
|---|---|---|
| `id` | str | 123 |
| `kind` | str | 123 |
| `phase` | str | 106 |
| `tactic` | str | 106 |
| `technique` | str | 106 |
| `how` | str | 106 |
| `mitre_id` | str | 106 |
| `tooling` | str | 106 |
| `limitations` | str | 106 |
| `grade` | str | 123 |
| `source` | str | 123 |
| `paired_mitigation` | NoneType/dict | 106 |
| `threat` | str | 17 |
| `detection` | str | 17 |
| `hardening` | str | 17 |

`kind` takes two values, `attack` (106) and `mitigation` (17).
The two record shapes differ:

- **attack rows** carry `mitre_id`, `phase`, `tactic`, `technique`, `how`, `tooling`,
  `limitations`, a long-form `grade`, `source`, and a nested `paired_mitigation` object
  with `threat`, `detection`, and `hardening` fields.
- **mitigation rows** carry `threat`, `detection`, and `hardening` directly.

All 106 attack rows carry `paired_mitigation`. Every
`mitre_id` value is a MITRE ATT&CK identifier (technique IDs such as `T1558.003` and
tactic IDs such as `TA0008` both appear).

## The `grade` field is not categorical

This matters for anyone filtering on it. On attack rows, `grade` holds **one of two
different things**:

- a short letter rating on 71 rows — `A` (62), `B` (1), `B+` (8);
- a long descriptive string (typically a high/medium/low judgement with its conditions) on
  35 rows.

Only 71 of 106 attack rows can be filtered into A/B/B+ buckets. The remaining
35 rows require parsing prose or dropping them. Treat `grade` as free text.

`phase` on attack rows:

| phase | rows |
|---|---|
| `Exploitation - Web` | 13 |
| `Initial Access` | 10 |
| `Evasion` | 7 |
| `Exploitation - AI/LLM` | 7 |
| `Vulnerability Discovery - Fuzzing` | 6 |
| `Exploitation - Memory Corruption` | 6 |
| `Lateral Movement - AD` | 6 |
| `Reconnaissance` | 5 |
| `Exfiltration` | 4 |
| `C2` | 4 |
| `Persistence` | 3 |
| `Credential Access` | 3 |
| `Privilege Escalation - Linux` | 3 |
| `Privilege Escalation - Windows` | 3 |
| `Reconnaissance - Fuzzing` | 3 |
| `Web Exploitation` | 2 |
| `Vulnerability Discovery - Symbolic Execution` | 2 |
| `Vulnerability Discovery - Binary Analysis` | 2 |
| `Lateral Movement / Credential Access` | 2 |
| `Privilege Escalation / Linux` | 1 |
| `Privilege Escalation / Active Directory` | 1 |
| `Command and Control` | 1 |
| `Web Exploitation / XSS` | 1 |
| `Web Exploitation / Injection` | 1 |
| `Vulnerability Discovery - OSINT / Recon` | 1 |
| `Exploitation - Binary` | 1 |
| `Vulnerability Discovery - Fuzzing / AI-Augmented` | 1 |
| `Evasion / Persistence` | 1 |
| `Exploitation - Mobile / Thick Client` | 1 |
| `Exploitation - Web / GraphQL` | 1 |
| `Reconnaissance / Initial Access` | 1 |
| `Exploitation - Web / Node.js` | 1 |
| `Evasion / Obfuscation` | 1 |
| `Exploitation - Physical / Hardware` | 1 |

The 17 mitigation rows carry no `phase`.

## Content scope

Attack techniques in the file span initial access, exploitation (including web and AI/LLM
targets), privilege escalation, lateral movement, credential access, and evasion. Each
record cites public sources (MITRE ATT&CK, public tooling repositories, public writeups)
in its `source` field, describes the technique, lists the public tooling that implements
it, and states its `limitations` — the conditions under which the technique fails or is
detected. The paired mitigation gives the corresponding `detection` and `hardening`
measures.

### Example record (excerpt)

```json
{"id": "ATK-005", "kind": "attack", "phase": "Privilege Escalation / Active Directory",
 "technique": "Kerberoasting — offline cracking of service account TGS tickets",
 "mitre_id": "T1558.003",
 "tooling": "Rubeus, Impacket GetUserSPNs.py, BloodHound/SharpHound, hashcat",
 "limitations": "AES-encrypted TGS tickets are significantly harder to crack than RC4 …
   Managed Service Accounts (gMSA) use 240-character random passwords — effectively
   uncrackable. Detection: Windows Event 4769 … with RC4 encryption type from
   non-service accounts is a strong indicator.",
 "source": "MITRE ATT&CK T1558.003 (public); Impacket GetUserSPNs.py (public OSS); …",
 "paired_mitigation": { "threat": "…", "detection": "…", "hardening": "…" }}
```

## Honest notes

- **Dual-use content.** This file describes offensive security techniques at a level of
  detail that includes tooling names and command shapes, alongside the detections and
  hardening measures that counter them. It is a public-source survey: every technique is
  attributed in `source` to publicly available material, and the record structure pairs
  each technique with its defensive counterpart. It is not exploit code — `how` describes
  technique, and no working payload is included.
- **Not independently verified.** The descriptions are compiled from public sources named
  in each row; no row carries a verification verdict, and the mitigation guidance has not
  been tested by the author.
- **`grade` is a judgement, not a measurement** — see above. The long-form values mix
  severity with conditional caveats in a single string.
- **`limitations` is empty on some rows.** Do not treat it as always-populated.
- **`paired_mitigation` is a cross-cutting pairing, not a topical one.** The mitigation
  paired with an attack is not necessarily the mitigation *for* that attack; pairings join
  across attack and mitigation pools. Do not read a row's `paired_mitigation` as its
  countermeasure.
- Example domains use reserved or fictional names (e.g. `domain.local`); no real
  credentials, hosts, or personal data are present. The file was scanned for secret
  material before staging.
- No held-out split is shipped; make your own.

## Files

- `attack_mitigation_corpus.jsonl` — 372607 B, 123 rows.
- `LICENSE` — CC-BY-4.0 (added at staging; the source directory carried no licence file).

## Licence

CC-BY-4.0. Free to use with attribution, including commercially.

Copyright 2026 Christopher Betances (catqualia.com)