CVAT CVE-2026-73220

Exploitation requires an attacker to hold a valid CVAT account with permission to create or edit annotation guides (CVSS PR:L - low-privilege authenticated user); unauthenticated exploitation is not possible. The vulnerability is scoped to the audio-task annotation guide feature within the audio workspace UI component; image or video task guides are not confirmed affected by this specific code path. A victim user must open the malicious guide in their browser (CVSS UI:P - passive user interaction), which in active annotation projects is a routine workflow step. Access to guides is typically limited to project members assigned to the task, narrowing the victim pool. No non-default configuration is required for the vulnerable rendering path to be active.

The CVSS 4.0 vector (AV:N/AC:L/AT:N/PR:L/UI:P) positions this as a network-exploitable, low-complexity stored XSS requiring low-privilege authentication and passive victim interaction, with high confidentiality and integrity impact (VC:H/VI:H) against the vulnerable system. Real-world risk is moderate-to-significant for collaborative annotation deployments: the stored nature means a single injected payload persists and can affect multiple victims over time without further attacker involvement, and in active annotation workflows, guide review is a routine, predictable action. No EPSS score was provided in the intelligence feed, so probabilistic exploitation likelihood cannot be quantitatively stated. The CVE is not listed in the CISA KEV catalog and no public proof-of-concept exploit has been identified, suggesting no confirmed active exploitation at time of analysis. The primary limiting factor is the PR:L requirement - an attacker must already hold a CVAT account with guide-editing permissions - which narrows the threat actor pool to malicious insiders or compromised annotator accounts.

An authenticated CVAT annotator with guide-editing access crafts an annotation guide for an audio task containing a Markdown-embedded JavaScript payload - for example, an inline HTML script block or a crafted image onerror handler. When a project manager or administrator navigates to the task and opens the guide for review, the unsanitized payload executes silently in their browser, issuing CVAT REST API requests under the victim's session token to exfiltrate annotation data or modify task assignments. No public proof-of-concept exploit has been identified, but the attack technique is well-established for stored XSS in Markdown renderers lacking sanitization.