Threat
GitLab Patches Critical 9.9 AI Gateway Flaw Allowing Command Execution on Self-Hosted Servers
Publication date not yet evaluated · Added · 2 reports thehackernews.com
Page published · Page updated
Coverage timeline
Why it matters
CVE-2026-90970 shows how an AI agent gateway's prompt-template sandbox can be weaponized into full remote command execution, making self-hosted LLM infrastructure a direct attack surface defenders must patch immediately.
GitLab patched CVE-2026-90970, a critical (CVSS 9.9) flaw in its self-hosted AI Gateway service powering GitLab Duo features. An authenticated user with Duo Agent Platform access could escape the prompt template sandbox via a specially crafted flow configuration, leading to arbitrary command execution on the AI Gateway. Fixes shipped in versions 19.2.4, 19.3.2, and 19.4.1; GitLab-hosted instances are already protected.
Summary
GitLab disclosed and patched a critical vulnerability tracked as CVE-2026-90970 in its AI Gateway service, which provides access to GitLab Duo AI features. The flaw could allow an authenticated user with Duo Agent Platform access to escape the prompt template sandbox and execute arbitrary commands on vulnerable self-hosted AI Gateway instances.[0][1]
The vulnerability, rated critical at CVSS 9.9, affects only GitLab Self-Hosted AI Gateway deployments; customers using the GitLab-hosted AI Gateway are already protected. GitLab released fixed versions 19.2.4, 19.3.2, and 19.4.1 and urged all self-managed customers running self-hosted AI Gateway to upgrade immediately.[0][1]
There is no evidence in the reporting that CVE-2026-90970 has been exploited in the wild; GitLab conducted targeted pre-disclosure outreach to self-hosted customers. A separate, unrelated GitLab path traversal flaw (CVE-2026-85706) was recently added to CISA's actively exploited catalog.[0][7]
Disclosure timeline
| Date | Event |
|---|---|
| 2026-10-02 | GitLab published a Friday advisory warning customers to immediately patch the critical CVE-2026-90970 AI Gateway RCE flaw and released fixed versions 19.2.4, 19.3.2, and 19.4.1.[0] |
| Before disclosure | GitLab conducted targeted outreach to Self-Hosted AI Gateway customers with upgrade guidance prior to the public release post.[0] |
How it works
The vulnerability is an improper neutralization weakness (CWE-1336). Under certain conditions, an authenticated user with Duo Agent Platform access can submit a specially crafted flow configuration that escapes the prompt template sandbox, resulting in arbitrary command execution on the AI Gateway.[0]
Exploitation requires basic privileges plus Duo Agent Platform access; successful abuse yields remote command execution on unpatched self-hosted AI Gateway instances.[0][1]
Affected versions and patch status
| Product | Affected | Patch status |
|---|---|---|
| GitLab Self-Hosted AI Gateway (deployed via GitLab Duo Self-Hosted on GitLab Self-Managed) | Self-hosted AI Gateway instances prior to the fixed releases | Patched in versions 19.2.4, 19.3.2, and 19.4.1[0] |
| GitLab-hosted AI Gateway (used by GitLab.com, GitLab Self-Managed, GitLab Dedicated) | Not vulnerable / already remediated by GitLab | No customer action required[0] |
Key takeaways
- CVE-2026-90970 demonstrates that AI-native service components such as prompt template sandboxes represent a new RCE attack surface when sandbox escapes via crafted configurations are possible.[0]
- The exposure is limited to self-hosted deployments, and GitLab's pre-disclosure outreach plus multi-branch patching (19.2.4, 19.3.2, 19.4.1) means defenders should prioritize confirming their gateway hosting model and applying the fix.[0]
- Although no in-the-wild exploitation of CVE-2026-90970 is reported, GitLab products have a history of actively exploited flaws (including the recent CVE-2026-85706 added to CISA KEV), underscoring the need for rapid patching.[0][7]
Defensive actions
- Upgrade GitLab Self-Hosted AI Gateway to version 19.2.4, 19.3.2, or 19.4.1 immediately.: These releases contain the critical security fix for CVE-2026-90970 and GitLab strongly recommends all self-managed customers with self-hosted AI Gateway installations update at once.[0]
- Restrict and review Duo Agent Platform access for authenticated users.: Exploitation requires an authenticated user with Duo Agent Platform access to submit a crafted flow configuration, so limiting that access reduces exposure on unpatched instances.[0]
- Verify whether your deployment relies on the self-hosted AI Gateway versus the GitLab-hosted gateway.: Only self-hosted AI Gateway deployments are vulnerable; GitLab-hosted AI Gateway users require no action.[0]