character-swap-h3

Warn

Audited by Gen Agent Trust Hub on Nov 4, 2026

Risk Level: MEDIUMDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [DYNAMIC_EXECUTION]: The test files assets/comfyui-h3-oneclick-swap/test_painted_area.py and assets/comfyui-h3-oneclick-swap/test_scene.py use ast.parse and exec(compile(...)) to dynamically execute specific parts of the skill's own __init__.py code for unit testing. While this is used for development verification, the use of exec on compiled source code constitutes dynamic execution of local data.
  • [INDIRECT_PROMPT_INJECTION]: The skill implements a workflow that relies on external vision-language model outputs to generate video generation prompts, creating an attack surface for indirect prompt injection.
  • Ingestion points: The nodes H3TrackRoles and H3FillShots process output from vision-language models (e.g., Qwen3-VL).
  • Boundary markers: The prompt templates defined in assets/comfyui-h3-oneclick-swap/__init__.py include instructions to the vision model to limit output to specific formats and subject identifiers.
  • Capability inventory: The skill facilitates full video generation through a ComfyUI server, involving file writes and model execution.
  • Sanitization: The skill includes neutral_shots and neutral_words functions in assets/comfyui-h3-oneclick-swap/__init__.py that filter the model output, removing gendered language and specific clothing descriptions to normalize the input before it is used in the final H3 generation prompt.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 4, 2026, 08:11 PM