kermt-finetune

Pass

Audited by Gen Agent Trust Hub on Aug 13, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes shell scripts and Python code within a local Docker container environment (kermt:latest) to manage system checks, data validation, and model training. It includes specific instructions for quoting target names to prevent shell metacharacter injection.- [EXTERNAL_DOWNLOADS]: Fetches official pretrained model bundles from Hugging Face. This process is protected by an explicit user consent gate or a dedicated opt-in flag, ensuring no remote content is downloaded without authorization.- [INDIRECT_PROMPT_INJECTION]: The skill has a data ingestion surface that processes user-supplied CSV files and model checkpoints.
  • Ingestion points: File paths provided via --csv, --ckpt, --val-csv, and --test-csv in SKILL.md.
  • Boundary markers: Implements a mandatory targets confirmation gate (Step 7) to verify training tasks before execution.
  • Capability inventory: Capabilities are limited to subprocess execution and file system access within the defined container scope.
  • Sanitization: Input data is sanitized and validated by specialized scripts (check_checkpoint.py and check_data.py) before being used in the training pipeline.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 13, 2026, 11:11 AM
Security Audit — agent-trust-hub — kermt-finetune