OpenClaw Vulnerability Exposes How an Open-Source AI Agent Can Be Hijacked

Researchers found a vulnerability chain in the OpenClaw agent whereby a malicious website's JavaScript can open a WebSocket to the locally bound gateway, brute-force or guess credentials because localhost attempts are exempt from rate limiting, gain administrative access, pair devices, and command the agent to exfiltrate data or execute arbitrary shell commands; OpenClaw issued a patch (2026.2.25) within 24 hours and the report urges immediate updates, credential audits, visibility into local AI tooling, and governance for agent identities.