Critical
GitLab, GraphQL, Code Injection, DevSecOps, Supply Chain

GitLab CVE-2026-19478: Unauth GraphQL RCE, Now Exploited
GitLab CVE-2026-19478 (CVSS 9.4), an unauthenticated GraphQL flaw, is now actively exploited (watchTowr, Aug 2026). Affected versions, fix, and mitigations.
GitLab CVE-2026-19478 is a critical, unauthenticated code-injection flaw in the GraphQL API of self-managed GitLab Community and Enterprise Edition that lets a remote attacker modify or delete public projects and user data without any credentials. It carries a CVSS score of 9.4, is categorized as CWE-94 (improper control of code generation), and was fixed in an out-of-band release on August 17, 2026. GitLab.com was patched before disclosure; every self-managed instance on an affected version needs to upgrade now. Updated August 22, 2026: CVE-2026-19478 is now under active exploitation. On August 21, 2026, watchTowr reported in-the-wild exploitation against its honeypot network, within days of disclosure; EPSS has climbed to the 72nd percentile.
The flaw lives in how GitLab handled custom GraphQL directives. Under certain conditions, unresolved directive input was dispatched dynamically against server-side objects, letting an unauthenticated caller invoke state-changing methods through an ordinary POST to /api/graphql. No session token, no cookie, no CSRF prerequisite. The same release also fixed CVE-2026-19650, a CVSS 7.1 flaw that allowed GraphQL mutations via GET requests due to improper multiplex query validation. Both were reported through GitLab's HackerOne program.
What makes this more than a defacement risk is the supply-chain angle. Silently altering or deleting a public repository is a way to stage a downstream compromise: hijack the projects that CI/CD pipelines and package consumers depend on, and the blast radius extends well beyond the GitLab server itself. That is why the security community expected weaponization fast, since the payload is a single unauthenticated HTTP request that automated scanners can spray across every internet-facing instance.
Which GitLab versions are affected and fixed?
The vulnerability affects self-managed CE and EE across four release trains. Upgrade to the patched build on your branch immediately; if you run an older, unsupported branch, follow GitLab's upgrade-path tool to step through the required intermediate versions.
Affected branch | Fixed version |
|---|---|
18.2 up to 18.11.11 | 18.11.11 |
19.0 up to 19.0.8 | 19.0.8 |
19.1 up to 19.1.6 | 19.1.6 |
19.2 up to 19.2.4 | 19.2.4 |
At disclosure there was no confirmed in-the-wild exploitation, but that window closed fast. On August 21, 2026, watchTowr reported in-the-wild exploitation against its honeypot network, only days after the patch, a case its researchers tied to AI-assisted exploit reproduction compressing the disclosure-to-attack gap. EPSS has since climbed to the 72nd percentile, and the small patch diff made reverse-engineering straightforward.
What can defenders do if they cannot patch immediately?
The primary fix is the upgrade; treat everything else as a stopgap. If a maintenance window is not available, restrict unauthenticated access to /api/graphql at your reverse proxy or WAF, or limit the endpoint to trusted internal ranges, accepting that this can affect front-end functionality. Temporarily reducing instance and project visibility shrinks the unauthenticated attack surface. For detection, inspect production_json.log and api_json.log for unauthenticated POST requests to /api/graphql with abnormal execution times or unknown directive syntax, and review audit logs for unexpected deletions or changes to public projects.
Related guides: How to Secure Your CI/CD Pipeline (GitHub Actions) and What Is CVSS? CVSS Score and Metrics Explained
Our read
An unauthenticated path from the public internet to object mutation is the kind of finding that a point-in-time scan reports as "informational" right up until someone deletes a repository. GitLab sits at the center of the software supply chain, so the honest question is not "is our instance patched" but "can we prove no one probed /api/graphql before we patched, and that nothing was altered." That is a supervised-adversary and compliance-as-evidence problem: exercise the exact unauthenticated request an attacker would send, keep the logs that show what happened, and you convert a scary CVSS 9.4 into a controlled, verifiable exposure.
Reporting by SecurityWeek; CVSS and affected-version data per NVD and GitLab's advisory. Sources linked above.
Related: GhostSplice and malicious MCP servers and the LiteLLM PyPI supply-chain incident.