auto-review-loop

Pass

Audited by Gen Agent Trust Hub on Oct 27, 2026

Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill uses the Bash tool and codex exec to run experiments, evaluate code, and manage project state. These actions are central to its autonomous research function.
  • [DYNAMIC_EXECUTION]: It dynamically assembles and executes shell scripts and AI model prompts to handle the multi-phase loop. It uses mktemp and cat to safely construct these inputs.
  • [EXTERNAL_DOWNLOADS]: The skill uses curl to fetch publication metadata and BibTeX records from academic sources like dblp.org and doi.org during the documentation phase to avoid fabricating citations.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted project files and experiment results as input for the AI reviewer.
  • Ingestion points: Reads code, logs, and experiment results from the repository (e.g., review-stage/AUTO_REVIEW.md, findings.md).
  • Boundary markers: Uses clear delimiters like === SCOPE LIMITS === and --- to separate instructions from data in model prompts.
  • Capability inventory: Includes the ability to execute shell commands, write files, and invoke other AI agents.
  • Sanitization: Specifically instructs against splicing raw artifact data directly into shell commands and utilizes local helper scripts for logic verification (e.g., review_gate.py).
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 27, 2026, 03:07 AM