compare-results
Pass
Audited by Gen Agent Trust Hub on Aug 17, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill defines a logical and secure workflow for benchmarking and comparing model checkpoints. The analysis found no evidence of prompt injection, obfuscation, unauthorized network activity, or credential theft. The author (nvidia) uses established patterns for data processing within an evaluation pipeline.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests data from external evaluation logs and artifacts, which constitutes an indirect prompt injection surface.
- Ingestion points: Score artifacts, task lists, configs, and logs are fetched from MLflow or NEL run paths (SKILL.md, Workflow Step 3).
- Boundary markers: The instructions include explicit validation steps, such as confirming the run passed a verification check (Step 4) and using canonical score fields from pre-defined task recipes (Step 5).
- Capability inventory: The skill delegates complex operations to other components (evaluation and MLflow skills) rather than executing arbitrary code itself.
- Sanitization: The use of structured extraction recipes and score field mapping acts as a sanitization layer by ignoring arbitrary content in the ingested logs.
Audit Metadata