CVE-2026-25253
WebSocket Hijacking via Malicious Skill
What this means in plain English
A malicious skill can eavesdrop on and tamper with messages sent between your other agents and the AI models they're connected to. This means a bad skill could read your private conversations, modify instructions mid-flight, or impersonate other agents entirely.
Technical description
A vulnerability in OpenClaw's WebSocket message handling allows a malicious skill to intercept and inject messages into other agent sessions sharing the same WebSocket connection pool.
Affected versions
< 0.13.2
Fixed in: 0.13.2
Details
Are you affected?
Run `openclaw --version` in your terminal
If your version is below 0.13.2, you are affected
Check if you have any skills with WebSocket permissions
How to fix
Update OpenClaw: `npm update -g openclaw`
Verify version: `openclaw --version` shows 0.13.2 or higher
Audit installed skills with WebSocket access via GitOpenClaw scanner
Scan your skills for this vulnerability
Use GitOpenClaw to scan any skill, repo, or install command for patterns associated with CVE-2026-25253 and other known vulnerabilities. Free, no account required.
References
- https://github.com/openclaw-ai/openclaw/security/advisories/GHSA-xxxx-1