The Agent Skills Directory

[PROMPT_INJECTION]: The skill documents an evaluation pipeline that processes untrusted data, creating a surface for indirect prompt injection.
Ingestion points: The example script in SKILL.md interpolates case.input and actual (untrusted model output) into the judge prompt.
Boundary markers: The prompt template lacks delimiters or specific instructions to isolate untrusted content from the instructions.
Capability inventory: Frontmatter enables the Bash tool, and the script uses the anthropic library for network-based LLM calls.
Sanitization: The provided implementation does not validate or sanitize inputs before interpolation.

ai-evaluation