The Agent Skills Directory

[PROMPT_INJECTION]: The skill documentation and scripts contain phrases typically associated with prompt injection, such as 'IGNORE ALL PREVIOUS INSTRUCTIONS'. Analysis confirms these are used as detection targets for a defensive scanner designed to protect the agent from malicious content in manuscripts, rather than as instructions for the agent itself.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted PDF manuscripts, creating an injection surface. It mitigates this by implementing a deterministic security gate that scans for hidden text layers (white-on-white text, tiny fonts, etc.) and metadata-based steering, sanitizing the input before it is processed by the LLM.
[COMMAND_EXECUTION]: The skill utilizes a series of local Python scripts for security auditing and quality assurance, including scan_pdf_layers.py for PDF extraction, check_pdf_injection.py for security auditing, check_review_request_types.py for reviewer discipline, and check_self_improvement_claims.py for methodological validation.
[EXTERNAL_DOWNLOADS]: The skill requires the pymupdf library for manuscript analysis. No suspicious remote code execution or unauthorized external data exfiltration patterns were identified.

peer-review