The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a model-based grading framework that processes code changes and open-ended AI outputs to assess quality and correctness. This creates a surface for indirect prompt injection if the content being graded contains adversarial instructions meant to influence the grader's evaluation.
Ingestion points: The [MODEL GRADER PROMPT] section in SKILL.md ingests code changes and open-ended text outputs for analysis.
Boundary markers: The provided templates do not currently include explicit delimiters or instructions to the model to ignore potential instructions embedded within the analyzed content.
Capability inventory: The skill employs tools including Bash, Write, Edit, Grep, and Glob to interact with the file system and run development commands like npm test and npm run build.
Sanitization: No explicit sanitization or filtering logic is defined for the content passed to the model-based grader, though the skill encourages human review for security-sensitive evaluations.

eval-harness