style-transfer-training

Pass

Audited by Gen Agent Trust Hub on Sep 4, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from users, specifically image datasets and restyling prompts, which are then passed to external machine learning APIs. This presents an attack surface for indirect prompt injection where malicious instructions could be embedded within the data to influence agent behavior. \n
  • Ingestion points: User-provided image corpora, input photos, and restyling prompts referenced in SKILL.md. \n
  • Boundary markers: The instructions in references/evaluation-protocol.md mandate a human review of all outputs and a side-by-side comparison between the base model and the LoRA to detect unexpected behavior. \n
  • Capability inventory: The skill uses network-based API calls to the fal.ai platform for training and inference operations. \n
  • Sanitization: The skill provides a 'Curation Checklist' in SKILL.md that requires manual subject isolation and the aggressive culling of atypical images, which helps reduce the ingestion of malicious data. \n- [EXTERNAL_DOWNLOADS]: The skill utilizes the fal.ai platform, as well as Replicate and Modal, for model training and inference. These are well-known services in the artificial intelligence infrastructure space, and their integration is appropriate for the skill's stated purpose of image style transfer training. \n- [COMMAND_EXECUTION]: The skill references the use of local shell commands like find and md5sum in references/anti-patterns.md for dataset integrity verification. These are standard system utilities used here for diagnostic purposes.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 4, 2026, 06:55 AM
Security Audit — agent-trust-hub — style-transfer-training