ZAHRA585/skillguard-v2-dataset

🛡️ 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

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:

@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}},
}