ai-scientist-evaluator

Pass

Audited by Gen Agent Trust Hub on Apr 12, 2026

Risk Level: SAFE
Full Analysis
  • [COMMAND_EXECUTION]: The skill utilizes a local Python script scripts/aggregate_reviews.py to aggregate and rank evaluation results. The command execution is localized to the skill's own environment and is used to process structured JSON data produced during the evaluation workflow.
  • [INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest and evaluate untrusted data (AI scientist outputs such as notebooks, manuscripts, and reports), which presents an inherent surface for indirect prompt injection where the processed data might attempt to influence the evaluation verdict.
  • Ingestion points: Artifacts including notebooks, code, figures, and manuscripts are processed by the agent as described in SKILL.md.
  • Boundary markers: The skill instructs the agent to act as a 'skeptical reviewer panel' and apply 'hard gates' in SKILL.md, providing a cognitive boundary, though no explicit technical delimiters are defined for the input data.
  • Capability inventory: The agent can read local files and execute the internal scripts/aggregate_reviews.py script for data processing.
  • Sanitization: No explicit sanitization or escaping of the external scientific artifacts is mentioned before the agent processes them.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 12, 2026, 02:58 AM