novelty-matrix
Warn
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: MEDIUMDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DYNAMIC_EXECUTION]: The script dynamically resolves the repository root at runtime to facilitate the import of shared library modules.\n
- Evidence: In
scripts/run.py, the code traverses parent directories to locate anAGENTS.mdfile and adds the resulting path to the system path usingsys.path.insert(0, str(repo_root)). This enables the dynamic loading of modules from a computed file path.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input data from external files, creating a surface where malicious instructions could be embedded and subsequently processed by an agent.\n - Ingestion points: The script reads from
output/CLAIMS.mdandoutput/PAPER.mdwithin the provided workspace directory.\n - Boundary markers: While the script uses parsing functions like
parse_item_blocksandextract_related_worksto structure the data, it does not implement explicit delimiters or warnings to prevent an LLM from following instructions potentially hidden in the paper text.\n - Capability inventory: The skill is capable of performing file read and write operations on the local filesystem.\n
- Sanitization: The script extracts and transforms text into structured artifacts (TSV/Markdown) but does not apply sanitization or escaping intended to neutralize prompt injection for downstream agent tasks.
Audit Metadata