The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill functions by analyzing external SKILL.md files, which introduces a surface where malicious content in those files could attempt to manipulate the evaluation report output. (1) Ingestion points: The skill processes external skill definitions and packages provided by the user for review as described in the evaluation protocol. (2) Boundary markers: The instructions lack explicit delimitation or instructions to ignore embedded commands within the input text being analyzed. (3) Capability inventory: The skill is purely instructional and does not utilize any tools for network communication, file system modification, or command execution. (4) Sanitization: No sanitization methods are specified for handling the untrusted text input.

skill-judge