The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied manuscripts, abstracts, and repository metadata (README.md, HF cards). This ingestion surface could theoretically be used for indirect prompt injection if a manuscript contained malicious instructions designed to mislead the agent's audit output. However, the risk is minimal as the audit is performed using local, deterministic scripts with no privileged execution of the input text.\n
Ingestion points: Manuscript drafts, titles, abstracts, and metadata files (README.md, CITATION.cff, HF cards) provided as Markdown or YAML.\n
Boundary markers: The skill instructs the agent to produce a structured visible checklist, providing a natural boundary for the agent's output.\n
Capability inventory: The skill uses Bash, Read, and Edit tools to run bundled Python scripts (check_summary_box.py, batch_metadata_audit.py, validate_schema.py).\n
Sanitization: The validation scripts use standard libraries (re, json) for structural and pattern-based analysis rather than content execution.\n- [SAFE]: All bundled scripts (Python and Bash) are transparent, use only standard libraries, and perform local validation tasks. The skill references well-known scholarly and technical services such as GitHub, Zenodo, Hugging Face, and PubMed, which are consistent with the academic use case. No indicators of credential theft, remote code execution, or persistence were found.

academic-aio