Your LLM security, tested and proven compliant.
~31,000 defensive adversarial prompts (injection, jailbreak, data leakage, toxicity, excessive agency, misinformation…) to replay against YOUR models with free tools (garak, promptfoo). The resulting report is your robustness evidence for the EU AI Act Annex IV file. On-prem, nothing leaves your infrastructure.
What we added recently
A living feed: here is the coverage added to it, dated.
- +274 sondes
- +82 sondes
- +58 sondes
- +104 sondes
Eight LLM attack surfaces covered.
LLM01, direct/indirect injections, DAN, obfuscation, many-shot, Unicode smuggling.
LLM02, system-prompt extraction, secrets, PII, prior-turn leakage.
LLM05, inert XSS/SQLi/SSTI/exfil canaries that downstream code may mishandle.
LLM06, proxy unsafe tool calls (delete, email exfil, shell) without approval.
LLM09, false facts, hallucinated citations/packages, false-premise agreement.
LLM10, resource exhaustion / denial-of-wallet (loops, recursive expansion).
+ two bundles from reference benchmarks (permissive): harmful-content refusal and toxicity generation. Every prompt is labelled with its technique + OWASP category + MITRE ATLAS technique, with a coverage report. ~50% of the content is original (in-house generated, proprietary).
Compliance evidence, not just prompts.
Every prompt → OWASP LLM Top 10 → EU AI Act (Art. 15 / Annex IV) → MITRE ATLAS → NIST AI 600-1.
Formats ready for garak (NVIDIA) and promptfoo. One command, a pass/fail report on YOUR models.
You run it in-house. No prompt or model is sent to a third-party SaaS, essential for regulated orgs.
Permissive sources (garak, MIT/Apache benchmarks) deduplicated + half original content generated by us.
Payloads = harmless canaries (PWNED, reveal your prompt…). Real malware/CBRN/PII filtered at build.
New techniques and categories added as LLM threats evolve. One key, always current.
How do I run it?
The pack ships prompts as JSONL, ready promptfoo tests (YAML) and a garak probe (tc_redteam.py). With garak: `garak --target_type ollama --target_name <model> --probes tc_redteam.Injection`. With promptfoo: `promptfoo eval --tests bundles/<bundle>/promptfoo-tests.yaml`. Both tools are free and open-source; point them at your model or endpoint.
How does it help with the EU AI Act?
EU AI Act Art. 15 requires robustness against adversarial attacks (adversarial examples, model evasion), and Annex IV requires the matching technical documentation. Replaying this feed against your system produces a test report, a direct documentation artifact for your compliance file. We also map NIST AI 600-1 (MEASURE actions) and MITRE ATLAS. Note: prompt-based coverage targets LLM01/02/05/06/07/09/10; architectural risks (LLM03/04/08) are process controls, not prompts.
What license? Can I redistribute it?
The pack is a proprietary subscription product (EULA included). You may use it to test your own systems; you may not redistribute, resell or build a competing feed from it. The ThreatClaw-generated original content (~half) and the compilation are our property; permissive third-party components keep their own license (attributed in the pack).
Ready to prove your LLMs are robust?
Annual subscription. Instant key. On-prem. Cancel anytime.