comfyui
Fail
Audited by Gen Agent Trust Hub on Aug 28, 2026
Risk Level: HIGHEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill automates the installation of external software and packages.
scripts/comfyui_setup.shusespipxorpipto installcomfy-clifrom official registries.scripts/auto_fix_deps.pyfacilitates the installation of ComfyUI custom nodes and models from the Comfy registry or specified URLs. - [COMMAND_EXECUTION]: Several scripts execute system commands to manage the ComfyUI lifecycle.
scripts/hardware_check.pyruns commands likenvidia-smi,rocm-smi, and PowerShell scripts to detect GPU and system resources.scripts/comfyui_setup.shexecutes shell commands to install and launch the ComfyUI server. - [INDIRECT_PROMPT_INJECTION]: The skill processes external workflow JSON files which can contain instructions or data that influence agent behavior. As noted in the skill's own documentation, ComfyUI workflows can execute arbitrary code via custom nodes, representing a significant capability surface.
- Ingestion points: Workflow JSON files are ingested by
scripts/run_workflow.py,scripts/extract_schema.py, andscripts/check_deps.py. - Boundary markers: The documentation explicitly warns that 'Workflow JSON is arbitrary code' and has the same trust profile as
eval. - Capability inventory: The skill can install packages, download models, write files, and trigger workflow execution on a server.
- Sanitization: The skill implements
safe_path_joininscripts/_common.pyto prevent path-traversal attacks when downloading outputs and includes logic inscripts/_common.pyto strip sensitive headers likeX-API-Keyduring cross-host redirects. - [DYNAMIC_EXECUTION]:
scripts/auto_fix_deps.pyandscripts/run_workflow.pydynamically interact with the ComfyUI server to install nodes or execute tasks based on the contents of the processed workflow files. - [REMOTE_CODE_EXECUTION]: Automated scanners flagged instances of
curl | python3(e.g., inscripts/comfyui_setup.shandSKILL.md). Analysis confirms these are used with the-m json.toolflag to pretty-print JSON data from the local server, which is a benign data-processing pattern rather than execution of remote code.
Recommendations
- HIGH: Downloads and executes remote code from: http://127.0.0.1:$PORT/system_stats, http://127.0.0.1:8188/queue - DO NOT USE without thorough review
Audit Metadata