CVAT CVE-2026-73219

The attacker must be an authenticated CVAT user holding write access to at least one job within the affected deployment (PR:L per CVSS 4.0 vector - unauthenticated exploitation is not possible). The attacker must know or be able to enumerate the numeric task ID of the target task they wish to disrupt; because CVAT task IDs are sequential integers, any authenticated user with read access to any task can trivially enumerate other task IDs. The target task must have the automatic annotation (lambda) feature available in the deployment. Exploitation is bounded by: (1) authenticated access is required - no remote unauthenticated path exists; (2) the disruption is scoped to the targeted task's annotation slot, not the entire CVAT instance; (3) the effect is reversible - an administrator can cancel the malicious request to restore access. No non-default configuration is required beyond a standard CVAT deployment with lambda function support enabled.

Real-world risk is moderate-low. The CVSS 4.0 score of 5.3 with VA:L accurately reflects a targeted feature-level denial-of-service rather than systemic platform disruption - the attack can stall another team's automated annotation pipeline but does not compromise data confidentiality or integrity, and does not affect CVAT availability more broadly. EPSS data was not provided in the input; the vulnerability is absent from CISA KEV and no public exploit code has been identified, consistent with a low exploitation urgency signal. The attack requires the adversary to hold authenticated write access to at least one job in the same CVAT deployment, limiting the threat surface to insider actors or accounts compromised within the same organization. CVAT task IDs are sequential integers, making enumeration of target task IDs trivial for any authenticated user, which slightly elevates the practical ease of targeting. The 'Information Disclosure' tag present in the source intelligence appears to be mislabeled - the CVSS 4.0 vector (VC:N, VI:N, VA:L) correctly classifies this as a limited availability impact with no confidentiality dimension; defenders should not treat this as a data leakage risk.

An authenticated CVAT user with write access to Job A posts a request to the batch annotation endpoint specifying Job A's ID but the task ID of an inaccessible Task B - whose sequential integer ID they enumerate by probing the API. The system reserves Task B's single active annotation slot for the attacker's malformed request, causing all subsequent legitimate annotation requests from Task B's owner to be rejected until an administrator intervenes. The attacker can resubmit the blocking request repeatedly to sustain the denial of service, stalling another team's ML data pipeline indefinitely.