GitLab 19.6 adds an audit log for AI agents
GitLab 19.6 records what AI agents do through its MCP server. A Ruby flaw is patched, but several NGINX flaws remain open, so admins who run GitLab themselves have settings to enable and a workaround to apply.

What GitLab 19.6 now logs
GitLab has released version 19.6 of its DevSecOps platform, as reported by heise online. From this version, every action an AI agent performs through GitLab's own MCP server creates an audit event. MCP is the protocol through which agents call tools in other systems.
The events appear in the top-level group under AI > Governance. The audit function is available in all editions, including the Free tier. Premium and Ultimate customers also get governance for the tools of the GitLab MCP server.
Enforcing agent restrictions on the server
Since version 19.4, owners of a top-level group can set for each MCP tool whether GitLab always allows it, always asks first or always blocks it. Until now the client had to enforce a block set with Always Deny. In 19.6 the MCP server enforces it for every connected client, and an agent does not even see a blocked tool.
This server-side enforcement is still in beta and depends on two feature flags that are switched off by default. If you want the restriction to hold whichever client is in use, you have to turn those flags on deliberately. The sources do not name the flags, so look them up in the official release notes.
Two further changes touch agents. The /goal command lets the Duo Agentic Chat start a long-running task as a developer-flow session. Duo flows in the runner used to continue without isolation when no sandbox was available and only issued a warning. A new option now makes the job fail in that case.
Ruby flaw closed, NGINX still open
Self-hosting admins should move to 19.6.1, which appeared on 20 November together with 19.5.8 and 19.4.10. It replaces the bundled Ruby and closes CVE-2026-132825, a heap buffer overflow in case conversion. The Ruby team considers remote code execution possible in some scenarios and rates the flaw critical. Linux packages and container images contain the fix, and GitLab.com is already patched. Installations from source need Ruby 3.3.14 or 3.4.13 rebuilt.
The bundled NGINX remains vulnerable to the five nginx2root flaws published on 18 November. GitLab 19.6 raises NGINX to 1.31.7, which closes five older flaws. The fixes for the new ones arrive only in NGINX 1.30.7 and 1.31.9. The most dangerous, CVE-2026-123007, sits in the HTTP/2 code and needs no login. Distribution updates do not help, because GitLab ships its own NGINX under /opt/gitlab/embedded.
What to check and do now
- Update to 19.6.1, 19.5.8 or 19.4.10, depending on your branch, or rebuild Ruby if you install from source.
- Set nginx['http2_enabled'] = false in gitlab.rb and run gitlab-ctl reconfigure. This mitigates two of the five NGINX flaws; the others concern HTTP/3, the DNS resolver and the mail proxy.
- Review custom directives in nginx['custom_gitlab_server_config'] and nginx['custom_nginx_config']. NGINX 1.31.7 treats buffer sizes of zero as errors, and with proxy_ssl_verify on it checks IP certificates only against IP SANs, otherwise returning a 502.
- Check whether your rate limit for authenticated Git HTTP requests still suits clone-heavy pipelines, since job-token access now counts towards it.
- Test runner jobs if you use FF_CONCRETE, because scripts with Bash arrays may fail when the shell is set to sh.
- Plan for Rapid Diffs, as the feature flags for the classic diff view are removed.
Ultimate customers can also add a malware rule to approval policies. It requires approval for a merge request whenever dependency scanning reports a package from GitLab's malware advisories. The beta function Rotate and fix revokes a leaked personal access token and opens a merge request that removes it. Systems outside GitLab CI/CD still need a new token afterwards.
Running agents with a log and approval
An audit log shows what an agent did, and an approval step decides what it may do. Whatever tool you use, check that agents have narrow permissions, that blocks hold on the server and that writing actions wait for a person. Our own platform follows that pattern: agents with approval prepare tasks, and anything that writes to a system waits for a release.
souveraen.ai is offered hosted in Germany or on-premise, and its security page describes the details. For wider questions on running models in your own network, see the overview of local AI. Apply the GitLab patches first, since the open NGINX flaws are reachable without authentication.


