The Agent Skills Directory

[Indirect Prompt Injection] (LOW): The skill workflow involves identifying and reading external SKILL.md files to provide improvements, which introduces an attack surface for instructions embedded in the target data. 1. Ingestion points: Skill identification in Step 1 and technical analysis in Step 3. 2. Boundary markers: Absent; the workflow does not define delimiters for separating untrusted skill content from agent instructions. 3. Capability inventory: The skill utilizes file reading and editing tools to modify system files. 4. Sanitization: Absent; no validation or filtering of the target skill content is performed during analysis. Mitigation: The requirement for a user-confirmed Improvement Plan in Step 4 acts as a critical safety barrier against autonomous exploitation.

improving-skills