ml-training-experiments
Pass
Audited by Gen Agent Trust Hub on Sep 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill directs the agent to analyze repository-level configuration files, lockfiles, and project metadata (e.g.,
pyproject.toml,CUDAnotes) to determine execution parameters. This behavior establishes an ingestion surface for indirect prompt injection from potentially untrusted source code. 1. Ingestion points: Repository-pinned Python versions, lockfiles, dataset manifests, and experiment configurations as specified in SKILL.md. 2. Boundary markers: The instructions do not define specific delimiters or 'ignore' warnings for content extracted from these files. 3. Capability inventory: The skill enables the agent to plan and execute tasks involving training loops, performance profiling, and interaction with experiment tracking systems (e.g., MLflow, Weights & Biases). 4. Sanitization: The skill does not prescribe explicit sanitization or validation steps for the metadata ingested from the repository environment. - [SAFE]: The skill includes explicit security recommendations, such as 'Never store secrets or tokens in notebooks' and performing 'dependency/security checks' before promotion, which align with industry best practices and mitigate potential data exposure risks.
Audit Metadata