modelopt-compare-results
Pass
Audited by Gen Agent Trust Hub on Nov 13, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external data including MLflow artifacts, evaluation logs, and task-specific recipes (
recipes/tasks/<task>.md). This constitutes an ingestion surface where untrusted data could theoretically attempt to influence agent behavior through embedded instructions. - Ingestion points: MLflow run artifacts, task recipes, and evaluation logs.
- Boundary markers: The skill relies on specific 'Score Extraction' blocks in recipes but does not define explicit delimiters for logs.
- Capability inventory: Data retrieval, statistical calculation, and delegation to other model-optimization skills.
- Sanitization: No explicit sanitization of log content or artifact metadata is described beyond score extraction.
- [EXTERNAL_DOWNLOADS]: The skill workflow includes fetching score artifacts and logs from MLflow and recording source URLs for external baseline sanity checks. These operations are within the expected behavior for a model evaluation and comparison tool.
- [COMMAND_EXECUTION]: While the skill delegates evaluations to other skills, it does not directly invoke shell commands or execute arbitrary code. It recommends using validated statistical libraries like SciPy for delta calculations.
Audit Metadata