GitLab Customer Turns on AI Agents to Fix Vulnerabilities Across 136 Projects in a Week
Scanners find vulnerabilities faster than people can fix them, and the queue only grows. In GitLab 19.6, a security manager can enable an AI agent across a whole group of projects at once instead of scripting it project by project. The agent writes a fix and opens a merge request. One customer enrolled 136 projects this way in the first week. If your backlog is large, start with a single project and the cautious Conservative preset: human review remains mandatory.
AI-processed from GitLab Blog; edited by Hamidun News
Imagine a security scanner flags hundreds of vulnerabilities in your repositories, while your developers manage to close about ten a week. GitLab released version 19.6 to address exactly this problem.
In it, a security manager can enable AI agents across an entire group of projects with a single action. Previously, that required an API script for each project. The agent checks whether a finding is a false positive, writes a fix, opens a merge request, and runs the tests, leaving a human to review the finished patch.
Everything is enabled through an ATR profile (automated triage and remediation). According to GitLab, one customer applied these profiles to 136 projects in the first week after the API was opened. If you have a large backlog, this is a way to offload the routine work without waiving mandatory approvals.
Why the vulnerability queue keeps growing
GitLab's blog describes it this way: AI writes more code and finds more holes, but people are still the ones fixing them. As evidence, the company cites NIST data. By its account, the National Vulnerability Database (NVD) has already published twice as many CVE records (public vulnerability identifiers) in 2026 as in all of 2025. Since April, NVD has stopped enriching most records with descriptions.
The second problem is that "it can wait" is no longer a safe choice. GitLab points to PHP2Shell: in October, frontier models found a pair of unauthenticated remote code execution vulnerabilities in PHP that had been judged unexploitable back in 2012. If attackers also use AI, holes once considered theoretical can be turned into working exploits within hours.
How agentic triage works
GitLab Duo Security agents scan the codebase and every merge request. They look for logic errors that conventional scanners miss and filter out false positives before they reach a developer's queue. What remains is ranked by severity, exploitability, reachability, and "blast radius," meaning what a breach would affect. To tie findings to the real state of the system, the agents use the GitLab Orbit knowledge graph.
For a confirmed vulnerability, the agent writes a fix, opens a merge request, and starts a pipeline. If tests or scans fail, it reads the logs, adjusts the patch, and checks again. The developer receives a merge-ready change with an explanation and the agent's confidence estimate. Approvals stay the same as for ordinary code and, GitLab says, neither people nor agents can bypass them.
What's new in 19.6
The ATR profile covers false-positive checks, risk assessment, and remediation. There are three presets: Conservative, Standard, and Proactive. They determine which severity levels each step applies to and whether it runs automatically or only on demand. Managers now select projects by group or attribute in Security Inventory or on the security settings page, assign a profile, and apply it. Security Inventory then shows which agent is enabled where and how well scanners cover each project. Profiles are available in the Ultimate tier across all GitLab deployment options.
The second addition is Vulnerability Context Flow, currently in beta. It identifies the affected service, including inside a monorepo, and adds a risk table to a SAST (static application security testing) finding. The table shows whether exploitation requires authentication, whether the service handles sensitive data, and whether it is reachable by untrusted input. The same finding in an isolated internal service and in a public API with personal data no longer looks equally urgent.
Third is an MTTR (mean time to remediate) chart on the security dashboard. It shows weekly values for first-party code and third-party dependencies at the project, group, and organization level, along with the number of closed vulnerabilities. The calculation counts findings that were fixed or are no longer detected; dismissed findings are not counted. To keep old backlogs from skewing the trend, each week counts only vulnerabilities found in the 90 days before that week began. The chart is on by default in Ultimate.
Separately, GitLab opened the Business Logic Security Analyzer (beta) for self-service enablement on GitLab.com. It looks for broken ownership logic, missed permission checks, and multi-step abuse scenarios. To run it on Anthropic's Claude Mythos 5.2, you need to contact your account manager. On self-managed and Dedicated, BLSA access is granted on request.
What to keep in mind
This is vendor material, and it contains no independent measurements. The figure of 136 projects says only how widely the feature was adopted, not how many vulnerabilities were closed or how many patches were accepted without edits. The text gives no false-positive rate and no data on fix quality. Everything described requires Ultimate, and Vulnerability Context Flow and BLSA are still in beta.
The practical takeaway is simple. Start with one project that has a large backlog: apply the Conservative profile, see how the agent handles fixes, and turn on exploitation context to re-sort the remainder. After a few weeks, the MTTR chart will show whether the queue is shrinking. Mandatory review remains the safety net until trust in the agent has been earned.
- Nov 20, 2026Mistral releases Vibe desktop app with parallel agent tasks
- Nov 20, 2026Suno Opens a Public MCP Server: You Can Now Generate Music from Claude and Cursor
- Nov 20, 2026Sourcegraph Cuts Deep Search Prices by 29% and Code Finder Calls by 73%
- Nov 20, 2026Pedestal AI Emerges From Stealth With $4.7M: Its Agents Cut Failed Deliveries by 20%
Key Questions Answered
Why the vulnerability queue keeps growing?
GitLab's blog describes it this way: AI writes more code and finds more holes, but people are still the ones fixing them. As evidence, the company cites NIST data. By its account, the National Vulnerability Database (NVD) has already published twice as many CVE records (public vulnerability identifiers) in 2026 as in all of 2025. Since April, NVD has stopped enriching most records with descriptions. The second problem is that "it can wait" is no longer a safe choice.
How agentic triage works?
GitLab Duo Security agents scan the codebase and every merge request. They look for logic errors that conventional scanners miss and filter out false positives before they reach a developer's queue. What remains is ranked by severity, exploitability, reachability, and "blast radius," meaning what a breach would affect. To tie findings to the real state of the system, the agents use the GitLab Orbit knowledge graph. For a confirmed vulnerability, the agent writes a fix, opens a merge request, and starts a pipeline.
What's new in 19.6?
The ATR profile covers false-positive checks, risk assessment, and remediation. There are three presets: Conservative, Standard, and Proactive. They determine which severity levels each step applies to and whether it runs automatically or only on demand. Managers now select projects by group or attribute in Security Inventory or on the security settings page, assign a profile, and apply it. Security Inventory then shows which agent is enabled where and how well scanners cover each project.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.