The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/normalize_paper.py uses importlib.util to dynamically load scripts/xray_io.py at runtime. While the path is constructed relative to the script's location, dynamic loading of modules from computed paths is a significant technique that could be abused if the filesystem is compromised.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest and process academic papers from external sources including arXiv, AlphaXiv, and user-provided PDFs. This data is untrusted and could contain adversarial instructions intended to bias the literature synthesis, review outlines, or strategic analysis produced by the agent.
Ingestion points: scripts/normalize_paper.py (fetching from URLs and parsing HTML/BibTeX) and scripts/xray_io.py (extracting text from local PDF and text files).
Boundary markers: The instructions in SKILL.md and reference modes do not explicitly define strict boundary markers or 'ignore' instructions when interpolating paper content into analysis prompts.
Capability inventory: The skill has access to Bash for running Python scripts and Write capabilities for persisting JSON artifacts to the local workspace.
Sanitization: The scripts perform basic text normalization and HTML unescaping, but lack specific sanitization logic to detect or neutralize prompt injection attempts within paper abstracts or full text.

paper-workbench