OpenClaw Flaw Could Let Malicious Websites Hijack AI Agents

Summary
Oasis Security reported that malicious websites could exploit OpenClaw’s localhost gateway to guess passwords and gain agent administrator access. OpenClaw fixed the high-severity flaw; users should update to version 2026.2.25 or later.
Key points
- Oasis Security reported that attackers could target OpenClaw agents when developers visited malicious websites; installing an extension or interacting with the site was not required.
- Browser JavaScript could connect to the agent’s localhost WebSocket gateway despite browser cross-origin protections.
- The gateway exempted localhost connections from rate limiting, allowing hundreds of password guesses per second in Oasis’s lab testing.
- After guessing the password, an attacker could register a trusted device without a prompt and gain an administrator session.
- That access could expose configurations and logs, access connected nodes, and enable actions such as reading messages or executing commands, according to Oasis.
- OpenClaw’s security team addressed the issue within 24 hours of receiving the report and classified it as high severity; users are advised to update to version 2026.2.25 or later.
Article Details
- Vulnerability Types
- Cross-origin WebSocket connections to the localhost gateway
- Password brute-force enabled by a loopback rate-limiter exemption
- Automatic approval of localhost device pairings without a user prompt
- Severity
- High
- Exploitation Status
- reported
- Exploit Availability
- unknown
- Patch Status
- available
