The Agent Skills Directory

[PROMPT_INJECTION]: The skill documentation includes a pattern for using an 'LLM-as-judge' which is susceptible to indirect prompt injection. * Ingestion points: The code snippets in SKILL.md show that user-provided 'query', 'response', and 'context' are directly interpolated into the JUDGE_PROMPT. * Boundary markers: The prompt template lacks explicit delimiters or specific instructions to the judge to ignore control sequences within the variables, making it vulnerable to data that attempts to override the judge's criteria. * Capability inventory: The evaluation logic is used for scoring and statistical monitoring; the snippets do not show high-risk capabilities like code execution or network operations being driven by the judge's output. * Sanitization: The provided Python examples use simple string interpolation without evidence of escaping or input validation.

evaluation-harness