distill-skill-to-model

Pass

Audited by Gen Agent Trust Hub on Sep 4, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process entire 'Agent Skill' packages, including their instructions, scripts, and tool definitions. This creates an attack surface where a malicious source skill could attempt to influence the distillation process.
  • Ingestion points: The pipeline resolves and hashes SKILL.md, referenced instructions, scripts, and assets from external source directories.
  • Boundary markers: The plan specifies removing the 'teacher-only' SKILL.md payload from training records and utilizes a pinned runtime manifest with a sandbox image to isolate execution.
  • Capability inventory: The skill involves executing the full workflow and tools of the source skill package to generate training trajectories.
  • Sanitization: The methodology includes layered verification of tool-call syntax, semantic and final-state correctness checks, human auditing of samples, and deduplication of training data.- [EXTERNAL_DOWNLOADS]: The skill references external model weights and software libraries from well-known repositories.
  • Evidence: Mentions downloading the Qwen/Qwen3-4B-Instruct-2507 model from Hugging Face and the MLX-LM framework from Apple's GitHub repository.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 4, 2026, 07:29 PM
Security Audit — agent-trust-hub — distill-skill-to-model