# GitLab 19.2: AI Agents Clear the Security Backlog You Created > Source: > Published: 2026-07-27 17:08:35+00:00 The same AI coding tools that shipped your features faster also shipped your vulnerabilities faster. GitLab 19.2, released July 16, closes the loop — four features now put AI agents to work on the security debt that AI-assisted development created. Security Review Flow and Dependency Scanning Auto-Remediation enter public beta. GitLab Duo CLI and Custom Flows hit general availability. The governance model is explicit: agents do the labor, humans make the final call. This matters because the numbers are bleak. [Veracode’s spring 2026 research](https://www.veracode.com/blog/spring-2026-genai-code-security/) found that 48% of AI-generated code contains security vulnerabilities. Security debt now affects 82% of organizations, up from 74% a year ago. Security teams were already struggling to clear backlogs from human-written code — AI tools made the volume problem mathematically worse. GitLab’s answer: automate remediation under human governance, not instead of it. ## Security Review Flow: Intent-Based, Not Pattern-Based Traditional static analysis tools catch vulnerabilities by matching known bad patterns. Security Review Flow works differently — it reasons about what your code is supposed to do, then checks whether the implementation actually does that. Any mismatch between intent and execution is a candidate finding. To use it: assign the `Duo Security Review` service account as a reviewer on any merge request. The flow analyzes the diff and posts threaded comments at the exact lines where vulnerabilities occur. Each finding comes with a CWE classification, severity rating, and where possible, an inline fix suggestion you can apply without leaving the MR. What it catches that pattern-based SAST misses: - Broken object-level authorization (BOLA/IDOR) - Missing authorization on state-changing operations - Mass assignment vulnerabilities - Business logic errors - Race conditions The governance model is deliberate: Security Review Flow never approves a merge request on its own. A human always makes the final call, and every review leaves a full audit trail. This is the right design. AI catching authorization logic errors is useful — AI approving code changes autonomously is a different risk category entirely. GitLab is threading that needle correctly. Security Review Flow is currently in [public beta](https://docs.gitlab.com/user/duo_agent_platform/flows/foundational_flows/security_review/). ## Dependency Scanning Auto-Remediation: It Fixes the Break Too Bumping a dependency version is the easy part. The actual work is fixing the code that breaks when a major version changes its API. GitLab 19.2 handles both. When Dependency Scanning Auto-Remediation detects a vulnerable package, it opens a merge request with the fix. If that version bump breaks the build, agents iterate in the same MR to resolve the breaking changes. Two independently configurable capabilities: straightforward version bumps, and agentic breaking-change resolution for upgrades that fail the pipeline. Configuration lets you set severity thresholds (which vulnerabilities trigger auto-remediation) and version scope (how aggressively it upgrades). Every change stops at your existing approval gates — the feature works with your current review process, not around it. This feature graduated from experiment (introduced in GitLab 19.0 behind the `dependency_management_auto_remediation` flag) to [public beta](https://docs.gitlab.com/user/application_security/remediate/dependency_scanning_auto_remediation/) in 19.2. The incubation period matters: this is a feature that had real-world iteration before it hit beta, not a rushed launch. ## Duo CLI and Custom Flows: AI You Can Script Against GitLab Duo CLI reaches general availability in 19.2 across all deployment models — GitLab.com, Self-Managed, and Dedicated, on Premium and Ultimate tiers. The distinction that matters most: it works headlessly. You can run Duo inside scripts and CI jobs, not just interactively at a terminal prompt. That makes it automatable in a way IDE-based AI tools are not. On Self-Managed 19.2+, Duo CLI is on by default. Administrators get controls for model selection, tool approvals, and MCP server connections — which means organizations running private inference can wire Duo CLI to their own models. Custom Flows also hit GA in 19.2. These are event-triggered agentic workflows — fired by GitLab events like mentions, assignments, pipeline failures, or MR lifecycle transitions — that replace manual multi-step processes. The Fix CI/CD Pipeline Flow, now improved, classifies failures before acting and delivers targeted fixes. The AI Audit Event Report, currently in beta, logs AI-assisted actions as dedicated audit events for compliance teams. Custom Flows GA is the platform play. The four foundational flows GitLab ships are a starting point; the real value is teams building their own event-triggered automations on a now-stable API. ## What to Enable First If you are on Self-Managed, GitLab 19.2 requires PostgreSQL 17 — check your database version before upgrading if you are running packaged PostgreSQL 16. Priority order for what to enable: start with Dependency Scanning Auto-Remediation (most immediate ROI, clearest governance model, no new review habits required). Add Security Review Flow to high-risk MRs next — it is in beta but the human-always-approves constraint keeps it safe to experiment with. Duo CLI GA is worth enabling if your team has recurring CI/CD pipeline debugging pain. Custom Flows are a longer-term investment once the first two are running. GitLab’s 19.2 release is worth reading as a statement of direction beyond the individual features: the company is betting that the answer to AI-generated security debt is governed AI remediation, not more human reviewers. [The official announcement](https://about.gitlab.com/press/releases/2026-07-16-gitlab-19-2-brings-governed-agentic-automation/) and [InfoQ’s coverage](https://www.infoq.com/news/2026/07/gitlab-19-2-ai-agents/) of enterprise reception are both worth reading. That bet is coherent. Whether the execution holds at scale — across codebases far messier than controlled demos — is what public betas will answer.