The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documentation suggests the installation of several well-known Python packages including bibtexparser, requests, semanticscholar, and arxiv. These are standard, legitimate libraries used for parsing bibliographic data and interacting with academic metadata services.
[COMMAND_EXECUTION]: The skill includes reference Python scripts (verify-citations.py, format-checker.py, api-clients.py) for the agent or user to execute. These scripts perform bibliographic formatting checks and metadata validation. They use standard argument parsing and do not contain any hidden or unauthorized command execution patterns.
[DATA_EXFILTRATION]: Network activity in the provided scripts is strictly limited to querying established scholarly API endpoints (e.g., api.crossref.org, api.semanticscholar.org) for the purpose of paper verification. No evidence of sensitive data harvesting or unauthorized transmission was found.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided BibTeX and LaTeX files. While this constitutes an attack surface for processing untrusted data, the scripts utilize standard parsers (bibtexparser) and implement text normalization (regex filtering and lowercase conversion), which mitigates common injection risks. The primary output is a verification report, presenting a low risk for influencing agent behavior through malicious data.
[PROMPT_INJECTION]: The instructional content in the markdown files focuses on academic integrity and best practices. No bypass instructions or attempts to override the agent's core safety guidelines were detected.

reference-audit-guide