GitLab 19.6 turns agent actions taken through its MCP server into audit events
GitLab 19.6 turns actions that agents take through its MCP server into audit events, but the beta feature records them only if AI audit event recording, which is off by default, is switched on. GitLab’s own NGINX is still vulnerable to nginx2root flaws, heise online reports.
GitLab released version 19.6 of its DevSecOps platform on 19 November. The MCP server is GitLab’s interface through which AI agents such as Claude Code, Cursor and Codex read and change GitLab data. The server now creates an audit event for every tool call, showing the actor, action, target and time. Events are marked as coming from the MCP server, so agent actions can be told apart from human ones.
GitLab infers sessions itself, its documentation says. The MCP protocol has no session identifier, so GitLab treats calls from the same user, through the same OAuth application, to the same project or group as one session. A session ends after a 15-minute pause and lasts at most 30 minutes, so it may not match a single conversation with an agent.
GitLab takes the agent’s name from the name of the OAuth application, which it does not verify. The documentation advises using the name for attribution only, not for authorisation. If the user has no role in the target project or group, the event is recorded in the user’s default GitLab Duo namespace or, if there is none, in the user’s personal namespace. Owners of the target project or group then cannot see the event.
MCP logging is in beta. Events are recorded only if AI audit event recording is switched on, and it is off by default. The events appear on the Audit events tab of the top-level group’s AI > Governance page, and viewing them requires the Owner role for the group. The release notes say the feature is available on all tiers, but the documentation lists AI audit events under the Premium and Ultimate tiers.
Germany’s heise online reports that patch releases 19.6.1, 19.5.8 and 19.4.10, published on 20 November, replace the Ruby bundled with GitLab and fix the Ruby2Shell vulnerability (CVE-2026-132825). The Ruby team says the flaw can, in some circumstances, allow remote code execution, and rates it critical. Users who install from source must rebuild Ruby 3.3.14 or 3.4.13.
GitLab 19.6 upgrades its bundled NGINX to version 1.31.7, fixing five older vulnerabilities. This version is still vulnerable to five nginx2root flaws published on 18 November, heise says. The most serious, CVE-2026-123007, is in the HTTP/2 code and can be exploited without logging in. The Linux package and Docker image use HTTP/2 for HTTPS connections by default, and distribution updates do not help because GitLab ships its own NGINX. Until GitLab delivers a fix, heise advises adding the line nginx['http2_enabled'] = false to the gitlab.rb file and running gitlab-ctl reconfigure. This mitigates only two of the five flaws, however.
According to GitLab’s upgrade notes, the new NGINX may break settings that administrators have added themselves, because it rejects zero-sized buffers and checks IP address certificates more strictly.
Read more: How an agent works · Agent map