The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/generate_pdf.py script invokes system-level binaries pandoc and xelatex via subprocess.run. Although the script uses argument lists to mitigate shell injection, it empowers the agent to execute external software on the host. The SKILL.md file also refers to a scripts/generate_schematic.py for diagram generation, which is not present in the provided file list, indicating a dependency on external or missing components.\n- [EXTERNAL_DOWNLOADS]: The skill relies on external software (pandoc, xelatex) and the requests library. The scripts/verify_citations.py script performs network operations to doi.org and api.crossref.org to retrieve and validate citation metadata.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from academic databases (e.g., PubMed, bioRxiv, arXiv), creating a surface for indirect prompt injection.\n
Ingestion points: Research data such as paper titles and abstracts are ingested via JSON files and processed by scripts/search_databases.py.\n
Boundary markers: The instructions do not define clear delimiters or specific guidance for the model to disregard instructions potentially hidden in the scientific content.\n
Capability inventory: The skill has Bash access and the ability to execute system commands and network requests.\n
Sanitization: There is no evidence of filtering or sanitizing the scientific text for adversarial prompt patterns.

literature-review