Malicious agent "skills" are sneaking past marketplace safety scanners: why it is not safe to use

What happened

Between July 6 and 9, 2026, researchers disclosed a technique called SkillCloak. It hides harmful code inside agent "skills" (add-on tools that give an AI assistant new abilities) so the code slips past the automated scanners that marketplaces run before listing a skill. The researchers tested it against 8 widely used scanners and more than 1,600 real malicious skills. It escaped detection over 90% of the time, and it worked on skills built for both Claude Code and OpenAI Codex.

The same research pointed to a campaign called ClawHavoc that had already planted 300 or more malicious skills on a public marketplace. Those skills tried to steal saved browser passwords, SSH keys (the login keys computers use to connect to each other), and crypto wallets.

What it means for you

A green "scanned" or "safe" badge is not proof a skill is safe. A badge tells you a scan ran. It does not tell you the scan caught everything, and this research shows the scan often misses.

What to do instead

• Install skills only from a source that reviews and tests each one by hand, not by scanner alone.
• Favor skills with a named author and a track record.
• Give a new skill the least access it needs to do its job.
• Watch for odd behavior right after you install one.

AgentPod lists only skills its team has reviewed and tested, which lowers this risk but does not erase it.

Sources

• https://www.helpnetsecurity.com/2026/07/09/malicious-ai-agent-skills-scan/
• https://cybersecuritynews.com/agent-skill-malware-targets-claude-code-and-openai-codex/