Attackers with GitLab access may exploit a flaw in the AI Gateway to run arbitrary commands on the host system. GajIPS detects attempts to exploit the vulnerability.
Severity: HIGH · Priority: P2 — authenticated Jinja2 sandbox escape in GitLab AI Gateway, investigate on hit
This rule detects attempted exploitation of CVE-2026-90970, a critical sandbox escape vulnerability in the GitLab AI Gateway component disclosed October 2, 2026. The gateway is the service that powers GitLab Duo's AI features and connects a self-hosted GitLab instance to language model providers. The flaw lives in the Jinja2 prompt template engine that the gateway uses to process AI flow configurations: a logged-in user with Duo Agent Platform access can craft a flow configuration that escapes the Jinja2 sandbox and executes arbitrary commands on the gateway host.
A hit indicates an attempt to exploit this sandbox escape against an AI Gateway in your environment. Confirm whether the targeted host is a self-hosted AI Gateway installation (GitLab.com and GitLab Dedicated deployments are already patched by GitLab and do not require action). If it is a self-hosted gateway, treat the attempt as potentially successful if the gateway was running a vulnerable version — exploitation requires only a logged-in GitLab user with Duo Agent Platform access, which is a low-privilege starting point. Upgrade the gateway Docker image to version 19.2.4, 19.3.2, or 19.4.1 as the definitive fix. Review gateway process logs for unexpected subprocess or shell execution since the vulnerability window opened.