The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill inherently processes large volumes of untrusted data from academic publications and web searches. This creates a surface for indirect prompt injection where malicious instructions could be embedded in the research material.
Ingestion points: The bibliography_agent and source_verification_agent specifically fetch and read external content from the web and local files.
Boundary markers: The skill mitigates this by implementing a 'canonical:instruction-data-boundary' pattern across its instructions, directing the model to treat all external content strictly as data and never as instructions.
Sanitization: The source_verification_agent and ethics_review_agent provide continuous verification and cross-referencing of all data entering the system, providing defense-in-depth against malicious content.
[EXTERNAL_DOWNLOADS]: The pipeline interacts with established academic registries (arXiv, Crossref, OpenAlex, and Semantic Scholar) for bibliographic verification. These interactions follow clearly defined API protocols and utilize environment variables for secret management (e.g., S2_API_KEY), which is consistent with secure development best practices. These sources are considered well-known and trustworthy in the academic research context.
[COMMAND_EXECUTION]: The timeline_extraction_agent is instructed to execute the 'pdftotext' system utility to extract metadata from locally stored PDF files. This is a standard research workflow and the utility is used in a specific, limited capacity to extract publication dates and other temporal facts from the research corpus.

deep-research