CVE-2026-84810

claude-skill-antivirus, a tool that scans local Claude skill directories for malicious content, only analyzes the SKILL.md manifest and ignores executable artifacts such as Python source files, bytecode, and other files in the skill's scripts directory. An attacker can distribute a skill whose manifest looks benign but whose scripts folder carries malicious code, and the scanner returns a SAFE verdict with a 100/100 trust score without ever examining those files. A user who relies on the tool's verdict then installs and runs the skill, executing attacker-controlled code on their machine under a false sense of assurance, with a high integrity impact (potential tampering with local files or data). Anyone using claude-skill-antivirus to vet third-party or downloaded Claude skills is affected; the available data specifies no affected or fixed version numbers. Exploitation has not been observed: there is no public proof-of-concept, the flaw is not in CISA KEV, and EPSS assigns only a 0.3% probability of exploitation within 30 days.

What to do: Do not rely solely on the tool's SAFE/100-score verdict: manually review every file under a skill's scripts directory (Python source, compiled bytecode, and other artifacts) before installing, and prefer skills from trusted sources. Watch for an updated release of claude-skill-antivirus that analyzes non-manifest files; no fixed version is identified in the available data. If any skill was installed based on a clean scan, re-inspect its scripts directory for unexpected or obfuscated code.