A Snyk analysis of 10,000 popular AI-agent skills found that more than 65% showed behaviour the researchers considered unsafe, including risks involving credentials, sensitive systems and untrusted content. The report also identifies a small number of skills it describes as overtly malicious, making the places people install agent skills a practical security concern, not just a developer convenience.
Watch Desk analysis
What happened
Snyk says it assessed the 10,000 most popular skills from the skills.sh and ClawHub marketplaces. Its researchers report that some unsafe skills could expose agents to sensitive systems, credentials or untrusted third-party content. They also identified a small number of malicious skills using lookalike domains, obfuscated code and covert background binaries to compromise user environments.
The findings are in Snyk’s analysis. The supplied report excerpt does not give a breakdown of findings by marketplace or explain how many skills fell into each risk category.
Why it matters
Agent skills can connect software to tools, accounts and information that matter. A skill that looks useful may therefore carry more than the ordinary risk of installing a dodgy extension: it could affect what an agent can access or what instructions it follows. Snyk’s figures point to a broad screening problem across the sample it examined, though they should not be read as a precise risk rate for every skill available.
Our read
Treat a skill like software you are choosing to trust, not a harmless prompt with a tidy label. Check who made it, what permissions it needs and whether its code or behaviour can be inspected. Convenient marketplaces are no substitute for knowing what you are letting into an agent’s toolkit.
What to watch
- Whether Snyk publishes a marketplace-by-marketplace breakdown and more detail on its assessment method.
- Whether the marketplaces add stronger review, provenance or reporting controls.
- Whether developers make skill permissions and external connections easier to inspect.
Discussion spark: Should marketplaces vet AI-agent skills before listing them, or should users bear the responsibility for checking what they install?
Sources and evidence
- Re-Assessing the Security of the Agent Skill Eco System (6 November 2026, 00:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.