Executive brief
CVAT is an open-source tool used by teams to label video and images for artificial intelligence training. A security flaw allows users with basic permissions to embed malicious scripts into annotation guides. If another user views these guides, the script can execute in their browser, allowing the attacker to perform actions or access data with the victim's authority.
Technical details
A stored Cross-Site Scripting (XSS) vulnerability exists in CVAT due to improper neutralization of input in the annotation guide component. The root cause is attributed to insecure defaults in the '@uiw/react-md-editor' library, which fails to sanitize Markdown content by default. An attacker with low privileges (the ability to create or edit tasks) can inject malicious JavaScript into the annotation guide. When a victim views the guide, the script executes in their browser context, enabling the attacker to perform unauthorized actions or exfiltrate data using the victim's session. The issue is resolved in version 2.64.0 by implementing proper sanitization for the Markdown editor.
Affected products
- CVAT.ai CVAT 2.5.0 to 2.63.0
Timeline
- 2026-04-30: advisory: GitHub Security Advisory published
- 2026-05-13: disclosed: NVD publication date