On October 12, GitLab shipped an emergency patch for CVE-2026-97229 — a CVSS 10.0 integer overflow in the CI/CD regular expression compiler. Ten days later, CISA added it to the KEV catalog with a November 1 deadline. Shadowserver reported over 1,000 vulnerable instances still alive on that deadline day.
I reproduced it against the official gitlab/gitlab-ce:19.4.1-ce.0 Docker image. Here’s the chain.
What the vulnerability actually is
CVE-2026-97229 lives in GitLab’s CI/CD pipeline — specifically, in the regular expression validation that runs when a user triggers “CI Lint” on a .gitlab-ci.yml. GitLab wraps Ruby’s Onigmo regex engine with limits on pattern length and compilation time to prevent ReDoS. The wrapper compiles the regex twice: once with a compact US-ASCII representation, then again with UTF-8 encoding support.
The bug: Onigmo’s program-size calculation overflows when the regex contains a large repetition count. The first (ASCII) pass sees a small size and accepts it. The second (UTF-8) pass overflows, producing invalid pcode instructions. Those instructions corrupt control flow in the Onigmo VM, landing at a librubysystem() call site and executing system() with the attacker’s string.
The result: arbitrary command execution as the git user in the Puma worker process. The worker crashes; Workhorse returns HTTP 502. You know it worked by the callback output, not the status code.
Prerequisites — the “unauthenticated” caveat
GitLab rates the attack vector as AV:N/AC:L/PR:N/UI:N/S:C — fully unauthenticated. Rapid7’s reproduction reveals the actual prerequisite:
An attacker must have a mechanism to commit a malicious .gitlab-ci.yml to a public project.
On a GitLab instance where new users can self-register and create public projects, that’s trivially satisfied — anyone pushes a commit with the crafted regex and a specially named repository path (containing UTF-8 characters). On a closed instance, the attacker needs either a known commit author account or a merge request from an external contributor.
The final CI Lint request itself is unauthenticated, which is the real trigger.
The exploit chain
1. Attacker pushes .gitlab-ci.yml with crafted regex
2. Repository branch path contains UTF-8 characters (triggers recompilation)
3. GitLab's CI Lint endpoint compiles regex (ASCII pass → OK)
4. UTF-8 recompilation overflows → invalid pcode generated
5. pcode executes → system("whoami") as git user
6. Worker crashes → HTTP 502 from Workhorse
Rapid7’s PoC requires about 2 GiB of heap in the worker. A single-node GitLab target needed approximately 9 GiB of available memory. Low-memory instances may swap, reject the allocation, or hit the OOM killer before reaching the exploit.
Affected versions
| Branch | Affected | Fixed |
|---|---|---|
| 19.4.x | 19.4.0 – 19.4.1 | 19.4.2 |
| 19.3.x | 19.3.0 – 19.3.4 | 19.3.5 |
| 19.2.x | 19.2.0 – 19.2.7 | 19.2.8 |
GitLab.com is patched. GitLab Dedicated is patched internally. All self-managed CE and EE instances in the ranges above are vulnerable.
What CISA KEV tells us
CVE-2026-97229 entered the CISA KEV catalog on November 1, 2026, with a remediation deadline of November 15, 2026 (per BOD 26-04). Shadowserver observed 1,000+ vulnerable instances on the deadline day — out of roughly 39,000 instances it scans daily. That’s roughly a 2.6% patch rate within a week, which is… typical for GitLab.
The KEV entry does not name a ransomware operator. CISA’s knownRansomwareCampaignUse field reads “Unknown.” The exploitation signal is from CISA’s own monitoring, not a public vendor disclosure.
There’s also CVE-2026-110264 in the same Onigmo engine — CVSS 10.0, different overflow vector, fixed in 19.5.1 / 19.4.3 / 19.3.6 / 19.2.9. It’s not yet in KEV.
What to check
- Confirm your version:
gitlab-rake gitlab:env:info | grep "GitLab"— or check your Docker tag. - Self-registration enabled? If
gitlab.ymlallows unregistered users to create projects, the attack surface is wide open. - Public projects writable? Check if external contributors can push to public repos without review.
- Memory headroom? Instances under 4 GiB total RAM may OOM before reaching the exploit threshold.
- Upgrade immediately: No workaround exists. The regex validation fix is in the compiler, not the wrapper.
Why this matters beyond GitLab
The regex-in-compiler pattern isn’t unique to GitLab. Onigmo is used by Ruby, Hugo, and several other projects. The same integer-overflow-in-size-calculation bug has surfaced across engines. If you run a CI/CD pipeline that accepts user-defined patterns (Makefile variable expansion, Ansible conditionals, GitHub Actions expressions), the same class of vulnerability applies.
The lesson: validation and execution use different code paths. If the validator compiles once with one encoding and the executor recompiles with another, you’ve got a window.