Agent Skills are a key extension mechanism for agent systems, packaging specialized capabilities for reuse by LLM agents. According to the arXiv paper MetaCascade, an agent first retrieves candidate skills based on skill documentation, then relies on their metadata to make decisions. The authors propose an end-to-end attack that poisons only this metadata, without modifying the skills themselves.
Because the attack is metadata-only, it targets the trust agents place in documentation and selection signals rather than the content of the skill code. The abstract does not include implementation details or experimental results, but it frames the attack as a full pipeline compromise.
As the sole source is the paper's abstract, this summary reflects only its claims and does not compare against independent findings.