paper-corpus-rag
Pass
Audited by Gen Agent Trust Hub on Nov 8, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The SKILL.md documentation suggests running Python scripts (
scripts/paper_index.py) via the command line to index and query local folders. While these are local operations on the user's own data, the script performs file system reads and writes (creating a SQLite database). - [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and summarize external research papers (PDF, Markdown). This creates an attack surface where malicious instructions could be embedded in the papers (e.g., hidden text or metadata) to influence the LLM's answers or citations.
- Ingestion points: Files in the
papers/orpapers_md/folders viascripts/paper_index.py(read inread_docfunction). - Boundary markers: None implemented in the provided scripts; the skill relies on the LLM to follow the 'Answering protocol'.
- Capability inventory: Local file reading, SQLite database creation and writing, and CLI output.
- Sanitization: The script performs de-hyphenation and regex-based chunking but does not sanitize content against embedded instructions before it is provided to the agent.
Audit Metadata