Malicious agent skills slip past 'safety' scanners for Claude Code and Codex: why it is not safe to use

What happened

In early July 2026, researchers at a Hong Kong university published work called SkillCloak. It showed that harmful "agent skills" (small add-ons that AI coding tools like Claude Code and OpenAI Codex load) can be dressed up to slip past the automated scanners built to catch them. One method hid the harmful code until the skill ran, and it fooled every scanner tested more than 90 percent of the time. A second method rewrote the code and beat over 80 percent of the "static" scanners, which check code without running it. The best scanner's catch rate fell from about 99 percent to around 10 percent. The team tested eight popular scanners against more than 1,600 real malicious skills, on a marketplace that grew past 40,000 skills within months of the format's launch in late 2025.

What it means for you

A "scanned" or "passed security review" label on a skill is not proof that it is safe. A skill can carry that badge and still cause harm once you install it.

What to do instead

• Install skills only from sources you already trust, not because a badge says "scanned."
• Favor small, curated collections over giant open marketplaces.
• Think before you let a skill reach your files, passwords, or accounts.

AgentPod lists only skills that people have reviewed and tested, and we say so plainly rather than promising perfect safety.

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/