literature-review

Pass

Audited by Gen Agent Trust Hub on Aug 28, 2026

Risk Level: SAFE
Full Analysis
  • [COMMAND_EXECUTION]: The skill documentation provides instructions for running a local Python linter (audit_review.py) and standard LaTeX build commands (latexmk). These commands are intended for the user or agent to execute within the project environment to ensure document integrity and generate PDFs.
  • [DATA_EXPOSURE]: The provided auditing script (scripts/audit_review.py) reads local project files including .tex, .bib, and .csv files. This access is scoped to the project directory to validate bibliography keys and evidence markers; no sensitive system paths or credentials are accessed.
  • [INDIRECT_PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection as it ingests and processes content from user-provided research files.
  • Ingestion points: Reads .tex, .bib, and source-ledger.csv files via the audit_review.py script.
  • Boundary markers: None; the script performs regex-based extraction of citations and identifiers.
  • Capability inventory: The skill includes instructions for shell-based auditing and LaTeX compilation but does not feature automated data writing or network exfiltration.
  • Sanitization: The script performs standard string matching and dictionary lookups without executing the content of the files.
  • [SAFE]: The skill instructions emphasize research integrity, explicitly directing the agent to avoid inventing sources or data, which aligns with safety best practices for LLM-based research agents.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 28, 2026, 12:50 AM