fine-tuning-expert
Pass
Audited by Gen Agent Trust Hub on Aug 27, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The file references/deployment-optimization.md provides code snippets utilizing subprocess.run to execute external Python scripts and binaries for llama.cpp conversion and quantization tasks. This functionality is standard for the described ML workflow but represents a direct command execution path.
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external training data in various formats (JSONL, Parquet, etc.). Ingestion points: load_custom_dataset function in references/dataset-preparation.md. Capability inventory: The skill provides logic for writing to the file system (save_pretrained, to_json) and executing system commands. Boundary markers: The code employs standard chat templates but relies on basic formatting. Sanitization: While a create_quality_filter is provided using regex to detect common AI signatures, it does not include specific sanitization against sophisticated prompt injection in training data.
Audit Metadata