The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill uses Python scripts to interact with public academic APIs, including api.crossref.org, api.semanticscholar.org, and export.arxiv.org. These are well-known academic services used to fetch citation metadata.
[COMMAND_EXECUTION]: The skill provides Python scripts (scripts/verify-citations.py, scripts/format-checker.py) that perform local file reading of .bib and .tex files and execute string-matching algorithms. These operations are restricted to the local workspace and are necessary for citation verification.
[DATA_EXFILTRATION]: While the skill performs network requests, they are limited to querying academic identifiers (DOIs, arXiv IDs, titles) against trusted academic repositories. There is no evidence of exfiltrating sensitive local files or environment variables.
[REMOTE_CODE_EXECUTION]: The skill requires standard Python libraries (bibtexparser, requests, semanticscholar, arxiv) for its operations. It does not use eval(), exec(), or download-and-execute patterns for untrusted code.
[PROMPT_INJECTION]: The skill instructions and documentation focus on academic integrity and do not contain patterns designed to override agent safety filters or bypass system instructions.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided bibliographic data (.bib or .tex files). While these files contain untrusted external data, the processing logic is limited to metadata extraction and comparison, posing a low risk of influencing the agent's behavior through embedded instructions in paper titles or abstracts.

inno-reference-audit