Your Agent, Their Asset: A Real-World Safety Analysis of OpenClaw

Researchers conducted the first real-world safety evaluation of OpenClaw, a widely deployed AI agent with extensive system access, revealing significant security vulnerabilities. The study found that poisoning any single dimension of the agent's state increases attack success rates from 24.6% to 64-74%, with even the strongest defenses still vulnerable to 63.8% of attacks.

• →OpenClaw's broad system privileges create substantial attack surfaces that existing sandboxed evaluations fail to capture.
• →The CIK taxonomy (Capability, Identity, Knowledge) provides a framework for analyzing AI agent vulnerabilities across three persistent state dimensions.
• →Attack success rates triple when any single CIK dimension is compromised, affecting all tested AI models including GPT-5.4 and Claude Sonnet 4.5.
• →Current defense mechanisms prove inadequate, with the strongest protection still allowing 63.8% attack success rates.
• →The vulnerabilities appear inherent to the agent architecture itself, requiring systematic safeguards beyond current approaches.