---
language:
- en
license: apache-2.0
task_categories:
- text-classification
tags:
- security
- prompt-injection
- llm-security
- agent-skills
- agentic-ai
- roberta
- cybersecurity
- vulnerability-detection
- benchmark
dataset_info:
  features:
  - name: text
    dtype: string
  - name: label
    dtype: int64
  - name: source
    dtype: string
  - name: attack_type
    dtype: string
  splits:
  - name: train
    num_examples: 19880
  - name: validation
    num_examples: 4260
  - name: test
    num_examples: 4260
  - name: adversarial_200
    num_examples: 200
  - name: skillsbench_infile
    num_examples: 1541
size_categories:
- 10K<n<100K
---

# 🛡️ SkillGuard v2: Complete Dataset & Adversarial Benchmarks Suite

**SkillGuard v2** is a curated dataset and benchmarking suite designed for training, evaluating, and stress-testing machine learning systems that detect **hidden prompt injections, credential exfiltration, and privilege escalation payloads in LLM Agent Skills (`SKILL.md`)**.

- **Domain:** LLM Agent Security, Indirect Prompt Injection, Adversarial Robustness
- **Language:** English & Software Code (Markdown, Bash, Python, YAML, Swift)

---

## 📊 Dataset & Benchmark Splits Summary

This repository includes both the **core training/testing splits** and **two independent held-out evaluation benchmarks**:

| Split Name | Samples | Benign (0) | Malicious (1) | Description |
| :--- | :--- | :--- | :--- | :--- |
| **`train`** | 19,880 | 9,940 (50.0%) | 9,940 (50.0%) | Core training set combining SkillMD-138k and synthetic threats. |
| **`validation`** | 4,260 | 2,130 (50.0%) | 2,130 (50.0%) | Hyperparameter tuning and checkpoint selection split. |
| **`test`** | 4,260 | 2,130 (50.0%) | 2,130 (50.0%) | In-distribution held-out test split. |
| **`adversarial_200`** | 200 | 100 (50.0%) | 100 (50.0%) | **Adversarial Zero-Day Benchmark**: 12 brand-new, unseen attack patterns injected into held-out skills. |
| **`skillsbench_infile`** | 1,541 | 1,500 (97.3%) | 41 (2.7%) | **SkillsBench-1650 Benchmark**: External benchmark testing in-file prompt injections across 8 categories. |

---

## 🧬 Threat Taxonomy & Attack Vectors

### 1. Synthetic Training Taxonomy (8 Categories in `train` / `validation` / `test`)
- **Data Exfiltration:** Hidden `curl`/`wget`/`nc` commands stealing SSH keys, AWS credentials, or tokens.
- **Guardrail Override:** Authority deception (claiming automated pre-approval or disabling confirmation dialogs).
- **Encoded Payloads:** Base64/Hex/rot13 encoded shell commands executed via `eval` or Python `base64.b64decode`.
- **Instruction Reprogram:** Prompt hijacking overriding the agent's core identity and role instructions.
- **Obfuscated Shell:** Variable expansion and subshell nesting to disguise dangerous binaries.
- **Dependency Poisoning:** Coercing agents to install malicious packages from untrusted third-party registries.
- **Hidden Comments:** Embedding injection commands inside HTML comments (`<!-- ... -->`) or markdown metadata.
- **Persistence / Backdoors:** Appending malicious aliases or cronjobs to shell configuration files.

### 2. Unseen Zero-Day Vectors (Tested in `adversarial_200`)
Tests out-of-distribution generalization on novel attack phrasings:
1. Fake GDPR Article 17 Data Portability compliance exfiltration.
2. Open Source License verification telemetry.
3. Automated Security Scanner access verification manipulation.
4. Skill Registry auto-update override.
5. OpenTelemetry distributed tracing exfiltration.
6. Anonymized usage analytics tarball exfiltration.
7. Jira elevation override authorization deception.
8. Internal pentest privileged mode flags.
9. Encoded Base64 runtime dependency bootstrapping.
10. Encoded preflight handshake verification.
11. SOC-2 compliance audit telemetry mandates.
12. Verified Publisher digital signature verification hooks.

---

## 🚀 Quickstart: Loading Datasets & Benchmarks

```python
from datasets import load_dataset

# 1. Load the entire suite
dataset = load_dataset("ZAHRA585/skillguard-v2-dataset")

print(dataset)
# Output:
# DatasetDict({
#     train: Dataset({features: ['text', 'label', 'source', 'attack_type'], num_rows: 19880}),
#     validation: Dataset({features: ['text', 'label', 'source', 'attack_type'], num_rows: 4260}),
#     test: Dataset({features: ['text', 'label', 'source', 'attack_type'], num_rows: 4260}),
#     adversarial_200: Dataset({features: ['text', 'label', 'source', 'attack_type'], num_rows: 200}),
#     skillsbench_infile: Dataset({features: ['text', 'label', 'source', 'attack_type'], num_rows: 1541})
# })
```

---

## 📁 Schema

| Column | Type | Description |
| :--- | :--- | :--- |
| `text` | `string` | Full content of the skill file or injected prompt block. |
| `label` | `int64` | Binary label: `0` for Benign, `1` for Malicious. |
| `source` | `string` | Origin split or benchmark name (`skillmd_138k`, `synthetic_v2`, `custom_adversarial_injection`, `skillsbench_infile`, etc.). |
| `attack_type` | `string` | Specific threat category (e.g. `data_exfil`, `unseen_injection`, `plaintext_cmd`, `none`). |

---

## 📜 Citation

If you use this dataset, its adversarial benchmarks, or the SkillGuard v2 model, please cite:

```bibtex
@misc{marouf2026skillguard,
  author = {Marouf Zahra},
  title = {SkillGuard v2: Transformer-Based Detection of Semantic Prompt Injections in LLM Agent Skills Under Adversarial Distribution Shifts},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/ZAHRA585/skillguard-v2-dataset}},
}
```