ClawSafety
2 projects share this name
ClawSafety
2 independent teams tackled AI agent security from opposite angles — one measures how safe agents are, the other makes them safer. Same name, complementary missions.
ClawSafety (Benchmark)
by weibowen555Independent (academic research)
2
Safety benchmark for personal AI agents under realistic prompt injection. 120 adversarial test cases evaluating whether frontier LLMs remain safe when serving as agent backbones.
Key Features
- •120 adversarial test cases across 5 harm domains (DevOps, Finance, Healthcare, Legal, SWE)
- •3 attack vectors: skill injection, email injection, web injection
- •Tested Claude Sonnet 4.6, Gemini 2.5 Pro, GPT-5.1, DeepSeek V3, Kimi K2.5
- •Scaffold comparison: OpenClaw vs Nanobot vs NemoClaw
- •Key finding: chat-safe models comply 40-75% under agent prompt injection
Best for: Security researchers and LLM developers evaluating agent safety. If you build or deploy AI agents, this benchmark tells you how vulnerable they are.
benchmarkprompt-injectionadversarial-testing120-test-casesarXiv-paper