CVE-2026-73220

This vulnerability exposes a fundamental trust model failure in CVAT's annotation guide system. A guide author with standard account access can embed malicious JavaScript that executes when any collaborator opens the guide, inheriting the victim's full session privileges. This means an attacker doesn't need admin access — they just need to create a guide and wait for someone to view it, then issue arbitrary CVAT API calls as that user. The attack surface is a feature designed for documentation, not security: annotation guides help teams standardize workflows, but the developers assumed guide authors were trustworthy by default. The privilege gap between "can create helpful documentation" and "can hijack any viewer's session" is enormous, and no sanitization existed to bridge it. The fix in version 2.70.0 — adding rehype-sanitize to the MDEditor pipeline — was technically straightforward, but the real issue is why it wasn't there initially. User-generated guide content wasn't treated as requiring the same sanitization as other user inputs, even though it renders HTML in a browser context with full session access. For defenders: verify your CVAT deployment is at version 2.70.0 or later. The more important question is what other user-controlled rendering paths exist with the same implicit trust assumption. Any feature that renders Markdown or HTML from unvetted users in a session-aware context is a potential repeat of this failure. Audit your codebase for MDEditor usage and other Markdown rendering points, and add a review checkpoint requiring contributors to document trust assumptions whenever code renders user-controlled content in a browser context.

Reviewed through automated stages and approved by a human before publication.