omh-model-finetuning
Pass
Audited by Gen Agent Trust Hub on Nov 9, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: No malicious patterns, obfuscation, or unauthorized network operations were detected. The skill is purely instructional and planning-oriented.
- [SAFE]: The skill explicitly instructs the agent to scan for and remove sensitive information such as personal data and credentials from training datasets before use.
- [INDIRECT_PROMPT_INJECTION]: The skill outlines a process for ingesting external data (demonstrations and preference pairs) for fine-tuning purposes. While it handles untrusted data, the risk is minimal because the skill defines a planner role that does not execute code, perform training, or make network requests.
- Ingestion points: User-provided training demonstrations and preference data described in SKILL.md and references/finetuning-method.md.
- Boundary markers: No specific prompt boundary markers (e.g., delimiters) are defined for data interpolation.
- Capability inventory: The skill explicitly specifies that it trains nothing, calls no models, and runs no evals, which severely limits the impact of any injected instructions.
- Sanitization: Includes a specific checklist requirement to scan examples for personal data and credentials and mask them before any processing occurs.
Audit Metadata