Chrome is one of the most widely used software products in history. Google's browser and its open-source foundation, Chromium, serve roughly 3.5 billion active users. That figure represents more than half of the world's adult population. With an estimated 73 percent share of the global browser market, Chrome is the primary gateway to the internet for billions of people.

Because of that enormous reach, even a single unpatched vulnerability can create worldwide risk. The Chrome security team is responsible for protecting not just a product, but a central piece of digital infrastructure. For years, human engineers handled the task of hunting for flaws and sending out patches. Now Google is bringing AI agents into that work with dramatic results.

The scale of the challenge

Google tracks browser releases in what it calls milestones. Milestone 128 arrived in August 2024. For each milestone from M128 through M145, the team fixed roughly 50 security bugs. That was already a considerable workload. Milestone 146 jumped to about 80 fixes. Milestone 147 handled roughly 130. Milestone 148 closed out about 350. Then came M149 and M150, where Chrome fixed 1,072 security bugs in a 60-day stretch.

That number is staggering because it surpasses the total number of security bugs fixed across the prior 23 milestones combined. In other words, the pace went from roughly 50 bugs per monthly release to more than 500 per release in the latest pair of milestones.

Key facts from the security push

• Google used Gemini-based AI agents to find vulnerabilities across the entire Chrome codebase.
• The AI workflow uncovered a sandbox escape flaw that had been present in Chrome since 2013.
• Google estimates the AI-driven triage process saves hundreds of hours of developer time per month.
• Chrome is now preparing to ship security updates twice a week in response to escalating attack velocity.
• The company is exploring dynamic patching to reduce the number of browser restarts required for updates.

How AI agents caught a decade-old bug

One of the most striking examples of AI-assisted security work is a sandbox escape vulnerability that originated in Chrome around 2013. A sandbox escape tricks the browser into reading local files that should remain off-limits. If exploited, it could allow an attacker to access sensitive data from a user's machine.

For more than ten years, that flaw sat in the codebase. Human reviewers and traditional test suites missed it. The vulnerability was not a simple case of unsanitized input. It emerged from complex interaction patterns that only became visible when the AI modeled numerous execution paths and scenarios.

Gemini found the bug in early 2026. The discovery illustrates a key difference between traditional scanning and AI-based vulnerability discovery. Static analysis looks for known dangerous patterns. AI agents can simulate and explore interactions, moving through possibilities much like a chess engine evaluating millions of positions.

The four-stage triage pipeline

Finding bugs is only half the battle. The Chrome team also had to handle an explosion of potential issues without overwhelming human engineers. Google built a four-stage triage operation.

In the first stage, AI agents filter out false positives, duplicate reports, and submissions that do not describe real security vulnerabilities. This step is described as 'filtering out the noise'. In the second stage, agents reproduce the reported bug in virtual environments that match the browser and operating system configuration. If the bug is confirmed, the agent adds details such as stack traces to the report.

The third stage enriches the report with metadata and background information. The fourth stage routes the complete case file to the appropriate human developer. That final handoff means engineers spend their time fixing code instead of assembling paperwork. Google said the process saves hundreds of hours of developer time every month.

Fixer and critic agents

After triage comes the actual repair work. Google described a multiagent workflow designed to scale patch creation. A fixing agent writes code and produces a set of candidate fixes. A critic agent evaluates those candidates, picks the most suitable one, and supplies additional context for the human reviewer.

The two agents cycle through a review loop similar to a traditional code review. They check that the code is functional and conforms to style guidelines. Then a separate group of small test-writing agents constructs tests and runs them across all of Chrome's supported platforms. If any test fails, the result is sent back to the fixer and critic before a human sees the patch.

Google says this automated review and testing process saves weeks of developer time for each release cycle.

Attackers have AI too

The new AI workflow is not just about convenience. It is a direct response to a growing security arms race. Google's AI tools can find vulnerabilities rapidly, but the same type of technology is available to malicious actors. It is unlikely that the flaws found and fixed by Google are the only vulnerabilities being discovered by outside researchers and attackers.

The speed of exploitation has increased as well. Time between the release of a security patch and the emergence of an exploit is shrinking. Google is therefore working to shorten its delivery cadence. The company wants to move from monthly milestones to a two-week cycle for major releases and a weekly cycle for security updates. Even that may not be enough. Chrome is now pushing to ship two security releases every week.

User behavior is the last bottleneck

Security updates only work if users install them. Chrome has relied for years on silent background downloads and automatic install after a restart. That approach has a weakness: the time between downloading a fix and restarting the browser is a window of exposure.

Google is trying to close that gap without making updates intrusive. One strategy is dynamic patching, which would apply many fixes without requiring a full restart. Another is preserving local browser state so users can resume their sessions quickly after an update. The company is also looking for opportune moments to restart Chrome without disrupting work.

Forcing restarts is unpopular, but leaving browsers unpatched is riskier. Google has to balance user experience against the immediate threat of exploitation.

What the acceleration means

Google found and fixed more than 1,000 vulnerabilities in two months. The previous 23 milestones produced a similar total. That kind of acceleration changes the timeline of software security. Work that would have taken a year can now be completed in a month. A decade of bug discovery may be compressed into a single year of AI-assisted development.

This acceleration is both an opportunity and a danger. AI gives defenders the ability to identify and fix more bugs than ever before. It also raises the ceiling for attackers, who can weaponize the same techniques. The result is a faster loop of discovery, exploitation, patch development, and deployment.

For Chrome's billions of users, this means the security team must rely on AI not only to find bugs but also to triage, fix, and verify them before the bad guys can strike. Google's agent-based workflow is designed to do exactly that. The process is not a complete replacement for human judgment; it is a high-speed assistant that lets engineers focus on the most difficult problems.

The long-term success of Chrome's security strategy will depend on how well these AI systems keep pace with AI-enhanced attacks. If Google can continue to compress months of manual labor into days, the browser can remain safe for the majority of the world's internet users. The future of web security now rests on machines fighting machines, with human oversight guiding every critical move.

Source: ZDNET News