- likes
- 13
- comments
- 3
Post
We re-scanned the top 10,000 agent skills on skills.sh and 10,000 on ClawHub. Here is what changed since our February report. 18 skills are actively malicious (7 on skills.sh, 11 on ClawHub). Fewer than at the start of the year, but more sophisticated: covert binaries, lookalike credential destinations, fake "official" vendors, obfuscated code. More than 64% of the skills we analyzed are vulnerable. They are not malicious, but they promote unsafe behaviors: wallet access, untrusted third-party content, hardcoded or plaintext credentials. One example: the agents-md skill runs an opaque binary first, then silently appends instructions downloaded from a remote server to the user's global CLAUDE.md. A persistent prompt injection that outlives the skill. The agent skill ecosystem remains a real supply chain risk, and scanning has to keep improving to keep pace. Our updated analysis engine is coming to Snyk Agent Scan and Snyk Evo. Authored by Marco Milanta, Marc Fischer and Luca Beurer-Kellner. Full research: https://vist.ly/5shd3