Skills are folders of instructions and scripts that an agent loads on demand, and the open Agent Skills format lets one skill run in several agents. Because a skill is mostly natural language, malicious intent can hide in setup steps that tell the agent to fetch and run code. Large 2026 audits found vulnerable and malicious skills in public registries.
Read in this order
- ArticleEquipping agents for the real world with Agent Skills (Anthropic)Start here: the design of skills and why they load progressively.
- ArticleToxicSkills: malicious AI agent skills study (Snyk)Audit results on how many public skills were malicious.
- PaperDetecting and Understanding Malicious Agent Skills in the WildA paper that classifies how malicious skills hide their behavior.