The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes AI-generated images which represent untrusted external data. This creates a surface for indirect prompt injection where an adversarial image could attempt to influence the multimodal reviewer. The risk is mitigated by explicit boundary markers in the judge prompt, OCR-based string validation, and the requirement for a fresh, context-free agent for every review batch.
[DYNAMIC_EXECUTION]: The rubric includes instructions for deterministic pixel measurement using standard image processing libraries (PIL and numpy). These operations are restricted to local image analysis and do not involve executing untrusted code or unsafe dynamic loading.

static-qa