CVE-2026-68771 describes an unsafe deserialization vulnerability in ComfyUI v0.23.0's LoadTrainingDataset node. Attackers can upload malicious shard_*.pkl files through an unauthenticated POST /upload/image endpoint. Subsequently, by queuing a workflow graph referencing the uploaded file via POST /prompt, the application calls torch.load on the attacker-controlled pickle file. This triggers the __reduce__ method during deserialization, allowing arbitrary Python code execution with the privileges of the ComfyUI process user. The vulnerability requires no authentication and has a CVSS 4.0 score of 9.3, indicating critical severity. No patch or official remediation guidance is currently documented.

Successful exploitation allows unauthenticated remote attackers to execute arbitrary Python code on the host running ComfyUI, potentially leading to full system compromise under the ComfyUI process user context. This can result in unauthorized control, data manipulation, or further attacks within the environment. The vulnerability is critical due to its remote, unauthenticated nature and high impact on confidentiality, integrity, and availability.

Patch status is not yet confirmed — check the vendor advisory for current remediation guidance. Until a fix is available, restrict access to the POST /upload/image and POST /prompt endpoints to trusted users or networks. Consider disabling or limiting the LoadTrainingDataset node functionality if possible. Monitor for updates from Comfy-Org regarding official patches or mitigations.