ClawHavoc: Koi Reported 341 Malicious Skills on ClawHub

Koi Security’s February 1 ClawHavoc disclosure counted 341 malicious skill listings on ClawHub, with 335 tied to one campaign. The figure counts marketplace listings, not victims.

OpenClaw skills give agents instructions while users may grant access to local files, shells, and authenticated services. Reviewing a skill is therefore a permissions decision: its instructions can steer tools already available to the agent.

After early reports, OpenClaw announced VirusTotal screening. Unit 42 later found five skills that bypassed or lacked verdicts from ClawHub’s VirusTotal and ClawScan checks. Scans help, but they do not replace checking what a skill asks an agent to do.

For teams using agent skills:

• Check the publisher and source repository before installation
• Read setup documents and linked pages, not just the marketplace summary
• Limit shell and file access to what the task needs

A careful skill review looks much like dependency review: verify origin, inspect behavior, and restrict permissions.