Google's Chrome browser is the gateway to the internet for more than half of the world's adults. With roughly 3.5 billion active users, the stakes are enormous. A single unpatched vulnerability can expose personal data, financial information, and private communications on a global scale. For years, Google relied on traditional security research, bug bounties, and fuzzing to keep Chrome safe. But in 2026, the company made a dramatic shift: it turned to AI agents powered by Gemini to find and fix vulnerabilities at a scale never seen before.
According to Google's Chrome security team, the latest AI-driven workflow identified and patched 1,072 security bugs in just 60 days across two milestone releases. To put that in perspective, that number surpasses the total number of security bugs fixed across the prior 23 milestones combined. The achievement marks a turning point in how one of the world's most complex software projects is secured.
The scale of Chrome's security challenge
Chrome and its open-source foundation, Chromium, are among the most intricate software projects ever built. The browser must interpret billions of web pages, run hundreds of thousands of extensions, and handle a dizzying array of media formats, all while maintaining sandboxing, isolation, and user privacy. Every new feature, every line of code, introduces potential attack surfaces.
With approximately 73% of the global browser market share, Chrome is the primary way most people access online services. This dominance makes it an irresistible target for attackers, nation-states, and cybercriminals. A single zero-day exploit can be used to compromise millions of devices before a patch is issued. For decades, Google's security team has fought this battle with human analysts, automated fuzzing tools, and community researchers. But the sheer volume of code and the speed of modern exploit development have overwhelmed traditional methods.
The hockey-stick growth in bug discoveries
Google's milestone releases, which occur roughly once a month, show a dramatic surge in bug-fixing activity. Milestone 128, released in August 2024, saw about 50 security bugs addressed. For the next 17 milestones, that number hovered around 50 to 80 per release. Then came a sharp acceleration: M146 fixed about 80 bugs, M147 fixed 130, and M148 fixed a staggering 350. For M149 and M150 combined, Google fixed 1,072 bugs.
The growth curve resembles a hockey stick, and it's not because Chrome suddenly became more insecure. Rather, it reflects a breakthrough in vulnerability discovery. AI agents are now scanning the entire codebase, identifying subtle interaction patterns that human reviewers and traditional static analysis tools have missed. This includes logic flaws, race conditions, and sandbox escape paths that have existed for years.
One particularly striking example is a sandbox escape vulnerability that has been present in Chrome since 2013. This bug, which could trick the browser into reading local files, passed test suite after test suite for more than a decade. Gemini found it in early 2026. The fact that a vulnerability can remain hidden for 13 years highlights the limitations of manual review and the power of AI-driven analysis.
How Google's AI agents work
Google's approach is not a single AI model but a multi-agent framework built around Gemini. The system was designed to handle the entire vulnerability lifecycle: discovery, triage, fixing, testing, and reporting. This is a departure from earlier AI applications in security, which focused mainly on fuzzing.
Fuzzing involves feeding a program with unexpected or random inputs to trigger crashes. Google has used AI to improve fuzzing since 2023, significantly expanding test coverage. By 2024, specialized AI tools were being used for vulnerability analysis. In 2025, Google collaborated with DeepMind and Project Zero on Big Sleep, an AI vulnerability discovery agent that successfully found bugs in the V8 JavaScript engine and the graphics stack. Big Sleep proved that AI could identify deep, complex vulnerabilities that humans often overlook.
The new agentic harness, based on Gemini, goes even further. It treats vulnerability discovery as a game of chess, simulating billions of interaction scenarios across the entire Chrome codebase. The AI doesn't just look for code errors; it looks for patterns of behavior that could be exploited. This requires an understanding of Chrome's architecture, its sandboxing model, and the ways different components communicate.
The four-stage triage automation
Discovering vulnerabilities is only the first step. Once bugs are found, they must be validated, prioritized, and assigned to developers. Google has automated much of this process with a four-stage triage operation.
In Stage 1, AI agents filter incoming bug reports, removing duplicates and irrelevant noise. This is analogous to clearing out spam from an inbox. For every legitimate vulnerability, there are countless false positives and low-quality reports. By automating this stage, Google saves its human analysts hours of tedious work.
Stage 2 involves reproducing the bug. The AI spins up virtual environments that mimic the specific browser version, operating system, and configuration described in each report. It then attempts to trigger the vulnerability. If successful, it adds valuable context such as stack traces and memory snapshots. This step is extraordinarily time-consuming for humans; what might take a developer several hours takes the AI moments.
Stage 3 enriches the report with metadata. This includes severity scoring, affected components, potential exploit paths, and historical context. The AI builds a comprehensive dossier, much like a detective preparing a case file before handing it to an investigator. In Stage 4, the AI identifies the ideal human owner for the vulnerability and automatically assigns the report, complete with all the necessary information. Google estimates this entire workflow saves hundreds of hours of developer time per month.
AI agents that fix and critique code
Once vulnerabilities are triaged, the fixing process also involves AI. Google has developed a multi-agent workflow where a fixing agent writes candidate patches, and a critic agent evaluates them. The two agents engage in a cycle reminiscent of a traditional code review. The fixer proposes a change; the critic examines it for correctness, style, and potential side effects. This loop continues until both agents are satisfied.
After a fix is approved, a small army of test-writing agents takes over. These agents create targeted regression tests and execute them across all of Chrome's supported platforms. If a test fails, the issue is sent back to the fixer and critic for revision. This automated testing process means that many bugs are caught and resolved before any human developer ever sees the code.
Google says this workflow saves weeks of developer time per release. It also allows the company to scale its response to the rapidly increasing number of vulnerabilities being discovered. Without AI, fixing 1,072 bugs in 60 days would be practically impossible, even for a team as large as Google's.
The update dilemma: patching fast without annoying users
Finding and fixing bugs is only half the battle. Getting the fixes into users' hands is the other half. Google has long used silent background downloads, staging updates on disk and applying them when the browser restarts. But many users rarely restart Chrome, and the window between a patch being downloaded and a restart occurring can be long enough for an exploit to take hold.
To address this, Google is shifting to more frequent release cycles. It has already moved to a two-week cadence for major milestones and weekly security updates. Now, the company is pushing for two security releases each week. That's a lot of updates, and user frustration is a real concern.
Google is experimenting with dynamic patching, which would apply fixes without requiring a browser restart. It is also working on saving browser state more effectively, making restarts less disruptive. Another idea is to identify opportune moments to restart, such as when the browser is idle or when the user's battery is fully charged.
Some users will resist forced updates, but the security risk demands action. As AI accelerates both attack and defense, the speed of patching must keep pace. Google is walking a fine line between being intrusive and being safe.
The AI security arms race
The same AI tools that Google uses to defend Chrome are available to attackers. Malicious actors can deploy Gemini-like agents to hunt for vulnerabilities in Chrome and other widely used software. If Google's AI can find 1,072 bugs in two months, an adversary's AI could find a similar number. The difference is that attackers don't need to disclose their findings. They can exploit them quietly.
This represents an explosive arms race. Every defensive breakthrough is matched by an offensive countermeasure. Google's blog post highlights just how intense this competition has become. The company is not only using AI to find and fix bugs; it is also using AI to understand how attackers might use AI.
The implications are profound. Two years' worth of vulnerabilities discovered in two months means that the pace of coding and security work is accelerating by a factor of roughly twelve. A year of AI-enabled development and attack will produce as much change as twelve years did in the pre-AI era. This acceleration threatens to outpace human oversight, making AI-driven defenses not just helpful but essential.
Google is publishing detailed technical information about its AI workflows, which is unusual for a company that guards its security methods carefully. This transparency may be intended to encourage collaboration across the industry, as every browser vendor, operating system maker, and cloud provider faces the same challenge.
For the 3.5 billion people who use Chrome, the immediate effect is a more secure browser. But the long-term effect is a world where software maintenance is increasingly handled by autonomous agents. That brings both promise and risk. A hallucinating AI could produce a flawed fix that breaks core functionality, and if that fix is pushed to billions of users, the consequences would be catastrophic.
Google's efforts show a way forward, but they also reveal how fragile the balance has become. The browser is no longer just a piece of software; it is a battlefield where machines fight machines under the watchful eye of human engineers. The only certainty is that this war will continue to escalate, and the tools needed to wage it will only become more sophisticated.
Source: ZDNET News