PwnClaw

PwnClaw is an AI agent security testing platform that runs 112 real-world attacks against your AI agent to find vulnerabilities before attackers do.

It tests across 14 attack categories including prompt injection, jailbreaks, social engineering, data exfiltration, MCP poisoning, agency hijacking, and more. Each scan randomly selects up to 50 attacks and uses an AI judge to evaluate whether your agent was compromised.

After each scan, PwnClaw generates copy-paste fix instructions for every vulnerability found. In benchmarks, Gemini 3 Flash went from 87/100 to a perfect 100/100 score with just 5 fix instructions.

Features:

• No SDK or API keys required — just paste your agent's endpoint
• CI/CD API and GitHub Action for automated testing
• Benchmark percentiles comparing your agent against frontier models
• Markdown export for security documentation
• Free tier available (3 scans/month)

Repo on GitHub.