The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides instructions and code for evaluating untrusted model outputs. 1. Ingestion points: The pred.answer property is ingested in the metric function templates in SKILL.md and the Appendix. 2. Boundary markers: Not present; untrusted input is directly interpolated into judge prompts. 3. Capability inventory: No dangerous capabilities such as shell execution, file writes, or network requests are included in the judge implementation. 4. Sanitization: No explicit sanitization or filtering is applied to the untrusted input. The skill mitigates this through architectural recommendations like decomposing judges into narrow yes/no tasks and performing mandatory human calibration.
[DATA_EXFILTRATION]: The documentation file references/R1-source-evidence.md contains hardcoded absolute file paths from the author's local development environment (e.g., /Users/5imp1ex/Desktop/Skill-Workplace/...). These represent static documentation artifacts rather than active attempts to access or exfiltrate data from a user's local system.

agentsop-metric-design