# CVE-2026-90970

Authenticated sandbox escape in GitLab AI Gateway enabling remote command execution.

- Page: https://www.vulneraibility.com/cve/CVE-2026-90970
- CVSS: 9.9
- Known exploited: not in CISA KEV as of our last sync
- EPSS: 17.49% (percentile 97.2)
- Unvetted proof-of-concept code on GitHub: 1 GitHub proof of concept (first 2026-10-03)
- CISA SSVC: exploitation none, automatable no, technical impact total (assessed 2026-10-02)
- Weakness: CWE-1336
- Published: 2026-10-02

## Affected versions

- GitLab GitLab AI Gateway: affected >= 18.1.6 and < 19.2.4; >= 19.3 and < 19.3.2; >= 19.4 and < 19.4.1

A version outside these ranges is not a statement that it is unaffected.

## Impact

Authenticated user with Duo Agent access can escape prompt template sandbox to execute arbitrary commands on the AI Gateway (RCE, lateral movement).

## Action

Immediate action required

## Recommendation

Upgrade AI Gateway to 19.2.4+, 19.3.2+, or 19.4.1+; restrict Duo Agent access and audit logs.

## References

- [other] https://gitlab.com/gitlab-org/gitlab/-/work_items/628842

---
Summary, impact and action are written by the vulneraibility.com AI analysis pipeline from the CVE record, NVD, CISA and FIRST data; verify against the references.
