ComfyUI vulnerabilities

ComfyUI v0.23.0 contains an unsafe deserialization vulnerability in the LoadTrainingDataset node. An unauthenticated attacker can upload a crafted pickle file via the POST /upload/image endpoint and trigger its deserialization via POST /prompt, leading to arbitrary Python code execution through torch.load.

ComfyUI prior to version 0.28.0 is vulnerable in folder_paths.get_annotated_filepath and exists_annotated_filepath. They join workflow-controlled filenames to a base directory without containment check, allowing an unauthenticated attacker to craft a POST /prompt request using LoadImage or similar nodes to probe arbitrary host paths and exfiltrate image-format files through /view.

ComfyUI prior to version 0.28.0 is vulnerable to stored cross-site scripting (XSS) in the GET /userdata/{file} endpoint. It serves user-controlled HTML and SVG files with extension-derived content types, allowing script execution in the ComfyUI origin and access to API tokens, settings, workflows, and authenticated-equivalent API calls.

ComfyUI prior to version 0.28.0 is vulnerable in the get_model_preview function in app/model_manager.py. It joins an unrestricted filename to a selected model directory without containment check, allowing an unauthenticated attacker to use traversal, absolute paths, or an unbounded path_index to read image-decodable files and enumerate host paths.

ComfyUI prior to version 0.28.0 is vulnerable to stored cross-site scripting (XSS) in the /view endpoint. It serves uploaded SVG files inline because image/svg+xml and related content types were not handled as dangerous, allowing script execution in the ComfyUI origin.