train-style-model

Pass

Audited by Gen Agent Trust Hub on Jul 2, 2026

Risk Level: SAFEDATA_EXFILTRATIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
  • [DATA_EXFILTRATION]: The skill requires users to host image datasets at public URLs for the platform to fetch. This workflow introduces a risk of data exposure for the user's private imagery if the URL is not secured or is inadvertently made discoverable.\n- [EXTERNAL_DOWNLOADS]: The training process involves fetching dataset ZIP files from external URLs provided by the user or the agent.\n- [PROMPT_INJECTION]: The skill processes untrusted image data, which serves as an ingestion point for indirect prompt injection where visual patterns could be maliciously designed to bias model behavior.\n
  • Ingestion points: Dataset archives via the inputs.dataset parameter in SKILL.md.\n
  • Boundary markers: No explicit delimiters or warnings for image content within the ZIP file are specified.\n
  • Capability inventory: Asynchronous training jobs and subsequent image generation tasks.\n
  • Sanitization: No documented filtering or validation steps for the content of the image files.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 2, 2026, 04:08 PM
Security Audit — agent-trust-hub — train-style-model