Fine-Tuning Assistant

Pass

Audited by Gen Agent Trust Hub on Feb 17, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [Indirect Prompt Injection] (LOW): The skill provides code templates for creating instruction-tuning datasets that interpolate untrusted data fields into message blocks, which could be exploited if the source data contains malicious instructions.
  • Ingestion points: The create_instruction_example function in SKILL.md accepts raw_data as input for prompt construction.
  • Boundary markers: Absent; the code snippets do not include delimiters or 'ignore' instructions to isolate user-provided content.
  • Capability inventory: Low; the snippets describe data transformation and training configuration, though the resulting models would be used for generation tasks.
  • Sanitization: Absent; there is no evidence of input validation or escaping in the provided documentation examples.
  • [Safe Practices] (SAFE): The skill does not perform active network operations, file system modifications, or command execution. It uses standard, well-known machine learning libraries and placeholders for configuration values.
Audit Metadata
Risk Level
SAFE
Analyzed
Feb 17, 2026, 05:19 PM