fine-tuning-expert

Fail

Audited by Gen Agent Trust Hub on Sep 15, 2026

Risk Level: CRITICALDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [DYNAMIC_EXECUTION]: The skill documentation includes examples using trust_remote_code=True when loading models from Hugging Face (e.g., in references/deployment-optimization.md). This allows for the execution of arbitrary code defined in the model's repository.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external datasets (JSONL, Parquet) for fine-tuning. These datasets represent an attack surface for indirect prompt injection if the source data is untrusted.
  • Ingestion points: Dataset loading in SKILL.md and references/dataset-preparation.md.
  • Boundary markers: The provided examples do not use explicit delimiters or instructions to ignore embedded prompts in the training data.
  • Capability inventory: The skill has access to file system writes (checkpoints), network operations (Hugging Face Hub), and subprocess execution.
  • Sanitization: A quality filter is provided in references/dataset-preparation.md, but it focuses on content suitability rather than security sanitization.
  • [COMMAND_EXECUTION]: The export_to_gguf function in references/deployment-optimization.md uses subprocess.run to execute Python scripts and binaries for model quantization, with paths derived from the LLAMA_CPP_PATH environment variable.
  • [EXTERNAL_DOWNLOADS]: The skill demonstrates fetching models and datasets from well-known repositories like the Hugging Face Hub (e.g., meta-llama/Llama-3-8B), which is a standard practice in the field.
Recommendations
  • CRITICAL: 1 file(s) identified as malware by FileRep - DO NOT USE
  • Contains 2 malicious URL(s) - DO NOT USE
Audit Metadata
Risk Level
CRITICAL
Analyzed
Sep 15, 2026, 08:39 AM
Security Audit — agent-trust-hub — fine-tuning-expert