train-character-lora

Pass

Audited by Gen Agent Trust Hub on Sep 7, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied images and captions to prepare a dataset for training. 1. Ingestion points: User-provided image paths and caption strings are ingested via the train_prepare_dataset tool. 2. Boundary markers: The instructions lack explicit delimiters to separate user-provided captions from the agent's control flow. 3. Capability inventory: The skill triggers train_start which launches a detached Docker container for training operations. 4. Sanitization: No explicit description of caption validation or sanitization is present.
  • [EXTERNAL_DOWNLOADS]: The skill facilitates the download of FLUX.1-dev model weights from Hugging Face. This is a well-known service and the operation is central to the skill's purpose.
  • [COMMAND_EXECUTION]: The skill manages the building and running of GPU-accelerated Docker containers via specialized MCP tools (train_doctor, train_start). This behavior is required for the primary purpose of local model training.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 7, 2026, 01:58 PM
Security Audit — agent-trust-hub — train-character-lora