Executive brief
CVAT is an open-source tool for annotating images and videos used in computer vision projects. A user with write access to any job in the system can block automatic annotation for other users' tasks by submitting a specially crafted request with mismatched task and job IDs. This disrupts the annotation workflow for other team members and prevents legitimate automated processing from completing.
Technical details
The vulnerability exists in the RequestViewSet.create endpoint, which accepts batch automatic annotation requests without validating that the supplied task and job IDs are consistent (i.e., that the job belongs to the specified task). Since the task ID determines the single active request slot per task, an attacker with job write access can submit a request using another user's task ID paired with an accessible job ID. This blocks that task's automatic annotation slot, preventing legitimate requests for that task from being processed. The vulnerability affects CVAT versions 2.17.0 through 2.71.x; a similar check already existed for interactive requests but was missing for batch requests. The fix was merged in version 2.72.0 and adds validation to ensure task and job ID consistency before creating a batch request.
Affected products
- CVAT AI CVAT 2.17.0 to 2.71.x
Timeline
- 2026-08-11: disclosed
- 2026-07-29: patched