knowledge-distillation
Pass
Audited by Gen Agent Trust Hub on Oct 8, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill includes instructions to clone the LMOps repository from Microsoft's GitHub organization to access MiniLLM implementation resources.
- [INDIRECT_PROMPT_INJECTION]: The skill's distillation and evaluation workflows ingest external text data, creating a potential surface for indirect prompt injection.
- Ingestion points: Data is processed through a dataloader in the distillation training loops and a prompts list in the evaluation section of SKILL.md.
- Boundary markers: The provided code examples lack explicit delimiters or instructions for the agent to ignore embedded commands within the ingested training text.
- Capability inventory: The skill is capable of performing model training, parameter optimization, and writing model weights to the local file system.
- Sanitization: No evidence of input validation or text sanitization for the training datasets or synthetic teacher responses is present in the snippets.
Audit Metadata