The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: Automated scanners flagged several commands that pipe output from curl to python3. However, manual inspection confirms these commands fetch structured data (XML/JSON) from trusted scientific domains (e.g., arxiv.org, openalex.org) and pipe it into local, bundled Python scripts (e.g., scripts/arxiv_atom.py, scripts/jats_to_text.py) for parsing. No actual remote code is being executed, and the data sources are well-known technology and academic services.
[EXTERNAL_DOWNLOADS]: The skill makes network requests to 18 well-known scholarly APIs, including PubMed, arXiv, and Crossref. These sources are established scientific repositories and the skill uses them strictly for data retrieval.
[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted content such as paper titles, abstracts, and full text from external APIs. It includes explicit security instructions to the agent to treat this data as untrusted third-party content and avoid following any instructions that might be embedded within it.
[COMMAND_EXECUTION]: The skill uses curl via Bash to interact with scholarly APIs. It employs safe patterns such as --data-urlencode to prevent command injection from user input and pipes data to dedicated local scripts for processing.
[DATA_EXFILTRATION]: The skill handles several API keys (e.g., NCBI_API_KEY, S2_API_KEY). It includes a shared helper script (scripts/_common.py) that implements a robust URL redaction mechanism to prevent accidental exposure of these credentials in the provenance logs or agent output.

paper-lookup