GitLab has issued an urgent security update after a critical vulnerability was discovered in the GitLab AI Gateway. The flaw, identified as CVE-2026-90970, allows authenticated users to bypass sandbox protections and execute arbitrary commands on affected systems.
The issue exists in the prompt template processing logic used by custom flows within the Duo Agent Platform. Under certain conditions, an attacker with valid access can submit a specially crafted flow configuration that escapes the sandbox environment and runs operating system commands.
According to the advisory, the vulnerability received a CVSS v3.1 base score of 9.9 and is rated Critical. GitLab has published fixed releases for self-hosted environments: GitLab AI Gateway 19.4.1, 19.3.2, and 19.2.4. The company states that its own hosted AI Gateway service has already been patched.
Administrators of self-hosted GitLab instances are strongly advised to apply the updates without delay to prevent potential exploitation.


