OpenClaw LLM Agent Bypass: Silent Execution Approval Disabled

The National Vulnerability Database has detailed CVE-2026-41349, a high-severity vulnerability affecting OpenClaw versions prior to 2026.3.28. This flaw allows LLM agents to silently bypass execution approval mechanisms. Attackers can exploit this by manipulating the config.patch parameter, effectively disabling user consent for operations.

The CVSS score of 8.8 (HIGH) reflects the critical impact: remote attackers with low privileges can achieve high confidentiality, integrity, and availability impacts without user interaction. This is a direct bypass of security controls designed to prevent unauthorized actions by AI agents, making it a dangerous avenue for stealthy operations.

For defenders, this means a critical need to patch OpenClaw deployments immediately. Any system relying on agentic consent for security must assume a compromised posture if running vulnerable versions. The attacker’s calculus here is clear: leverage the growing reliance on LLM agents to bypass traditional human-in-the-loop security, executing malicious tasks under the radar.