comfyui

Pass

Audited by Gen Agent Trust Hub on Sep 12, 2026

Risk Level: SAFEPERSISTENCECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [PERSISTENCE]: The UPDATING.md file provides explicit instructions for users to set up automated update schedules using Windows Task Scheduler (schtasks) or Linux cron. These tasks execute the check_updates.py script weekly to synchronize the template library.
  • [COMMAND_EXECUTION]: The tools/check_updates.py script uses the Python subprocess module to execute system commands, specifically git pull and git rev-parse. These commands are used to maintain the local clone of the official ComfyUI template repository.
  • [INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of untrusted data through tools/fetch_workflow.py, which downloads user-shared workflow JSON files from comfy.org. These workflows are then executed by the agent via the comfy_client.py API wrapper. This creates a surface where a malicious workflow could contain instructions that influence the agent's behavior during the generation process.
  • Ingestion points: tools/fetch_workflow.py downloads JSON files from an external repository.
  • Boundary markers: The skill does not implement explicit boundary markers or 'ignore embedded instructions' warnings for downloaded workflow content.
  • Capability inventory: The skill can execute complex multi-step generative graphs and file system operations (image/video saves) via the ComfyUI API.
  • Sanitization: No explicit sanitization of the JSON structure or embedded metadata is performed before the agent processes the file.
  • [PROMPT_INJECTION]: The skill includes technical guidance in MODELS/image-open-more.md and MODELS.md regarding the use of 'abliterated' text encoders. These instructions explain how to swap standard encoders for modified versions designed to bypass safety-alignment refusals in generative models (such as Qwen and Krea). While the skill defines a policy for the agent to only use these when requested, it provides the exact method and resources required to circumvent model-level content restrictions.
  • [EXTERNAL_DOWNLOADS]: The skill performs several network operations to well-known services:
  • Fetches the official announcements RSS feed from blog.comfy.org.
  • Synchronizes the official template library from the Comfy-Org GitHub organization.
  • Downloads model weights from official huggingface.co repositories (e.g., Stability AI, Black Forest Labs, Lightricks) via curl as instructed in SKILL.md.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 12, 2026, 01:07 AM
Security Audit — agent-trust-hub — comfyui