AI bug detection has grown into agents that scan code, prove exploits, and draft automated code fixes, yet human developers still need to review, validate, and own every security change that reaches production.
The headline story in AI bug detection today is not that tools find more vulnerabilities—it is that AI security agents now scan code, confirm exploitability, and write automated code fixes, while developers remain accountable for what ships. Google’s preview of CodeMender, an agent that scans repositories for security flaws, proves they can be exploited, and generates patches for developer review, shows how far this has moved beyond static vulnerability detection tools. GitLab’s 19.2 release takes the same agent-first stance, targeting the security backlog created when AI coding tools produce more changes than teams can manually review. The critical point: every one of these systems is designed with a human checkpoint. They reduce toil and false positives, but they do not—and should not—remove the developer from the decision loop.
Google CodeMender: From passive scans to exploit-backed patches
CodeMender is Google’s answer to attackers who are already using AI to speed up exploitation, arguing that defenders need automation running at the same pace. Instead of flagging patterns and walking away, the agent works in three stages: it first scans code in languages such as C/C++, Go, Java, Python, Ruby, Rust, and TypeScript for issues that static tools often miss, including memory corruption, injection flaws, web security problems, cryptographic weaknesses, and insecure data handling. Then it attempts to build a proof-of-concept exploit and run it in a customer-controlled sandbox, cutting the noise that usually overwhelms security teams. Once a flaw is confirmed, CodeMender writes a patch and passes it to a judge model to catch regressions, delivering the result as a code diff that still requires manual approval before any repository is changed. This is automated remediation with guardrails, not auto-merge magic.
GitLab 19.2: AI security agents attack the AI paradox
GitLab is blunt about the problem it is trying to solve: coding agents now generate far more code, moving the bottleneck downstream to reviews and security. Version 19.2 of its DevSecOps platform adds agentic automation aimed squarely at that backlog. Dependency Scanning Auto-Remediation opens merge requests when a scan finds vulnerable dependencies and proposes safe versions; if an update breaks the build, an agent iterates within the same merge request until the pipeline passes, a flow GitLab calls Agentic Breaking Change Resolution. According to GitLab, a study of the Maven ecosystem found that vulnerabilities reach roughly 63 percent of latest releases through transitive dependencies, and about one in eight dependency updates introduces a breaking change. That is exactly the kind of drudgery AI security agents are well suited to clean up—reducing noisy, repetitive work without bypassing the maintainer who must still review and approve every change.
Logic flaws, custom flows, and the limits of autonomy
Dependency scanning is only part of the story. GitLab’s Duo Security Review Flow targets logic flaws that pattern-matching scanners routinely miss, including broken authorization checks, mass assignment, and race conditions. Teams assign a service account as a reviewer, and the agent leaves threaded comments with severity ratings and proposed fixes—but it never approves a merge request on its own; a person still makes the final call. This is a conscious design choice: AI security agents are advisors, not gatekeepers. GitLab Duo CLI brings those agents into the terminal with awareness of pipelines and configuration, while Custom Flows let teams define their own automations in YAML and trigger them on GitLab events. There is even an AI Audit Event Report in beta and new access controls for which agents can run and what they can reach, an acknowledgement that autonomous behavior needs traceability and limits to stay trustworthy.
Developers still own the review—and the fallout
The emerging pattern across these AI security agents is clear: they attack the widening gap between code generation and security review capacity, but they do not remove human responsibility. GitLab stresses that maintainers must review and approve every automated dependency fix before it merges, so the agent clears manual patching work without touching the governance checkpoint. Google’s CodeMender is built with the same constraint—its patches arrive as diffs, and no change reaches a repository without manual approval. AI can now scan, exploit, and propose repairs faster than any human-only process, and a commissioned study even claims organisations using GitLab’s Duo Agent Platform saw a 400 percent return on investment with payback in under six months. But when an AI-generated fix introduces a subtle bug, it is the development team that answers for it. That reality should shape how leaders adopt vulnerability detection tools: as power tools for security, not replacements for professional judgment.
