Google、AI エージェントで Chrome のセキュリティバグ 1,072 件を 60 日間で発見・修正

*Follow ZDNET: *Add us as a preferred source* on Google.*

ZDNET's key takeaways

• AI found a decade-old Chrome flaw humans repeatedly missed.
• Google said its AI workflow saves weeks of developer time.
• Chrome may need twice-weekly patches to keep billions safe.

Chrome and its open-source project, Chromium, are, by scale, among the most complex open-source projects on the planet. With some 73% of the global browser market share, there are about 3.5 billion active Chrome users. To put that in perspective, 3.5 billion is more than 10 times the entire population (adults and children) of the United States.

I couldn't find a good reference for the overall adult population, so I used math. Based on data for the overall population of the world and the number of children on the planet, there are a little under 6 billion adults on Earth. The math shows us that Chrome is the primary way more than half of us earthlings access the internet.

Also: Claude AI shared chats indexed by Google - see if your conversations were exposed

The stakes, therefore, are incredibly high for Chrome's developers. Just letting one security flaw loose into the wild immediately puts half the adult humans currently alive at risk.

Google clearly takes this responsibility quite seriously. Last week, the Chrome security team released a blog post detailing how it is using AI to "improve vulnerability discovery, triage, and patching."

Also: This powerful Gemini setting made my AI results way more personal and accurate

The scope, the challenge, how AI is helping, and how AI is also exposing overwhelming numbers of new vulnerabilities make for quite the story, one worthy of deconstruction.

The whole story in one chart

Google called its releases "milestones," which occur about once a month. The chart below begins with Milestone 128, which dropped on Aug. 20, 2024. For each milestone up to M145, Google fixed roughly 50 bugs. Then M146 jumped to about 80 bugs. M147 fixed about 130 bugs. M148 fixed about 350 bugs.

*The chart shows a hockey-stick growth curve. Image: Google*

Month by month, the number of bugs found and fixed doubled. Then, for M149 and M150, Google fixed 1,072 bugs, "surpassing the total number of security bugs fixed across the prior 23 milestones combined."

There are ... implications

The logistics of this are mind-blowing. With Google issuing major new releases on a monthly basis, the potential for causing disruption worldwide is fairly high. If, instead of fixing a bug, Google were to introduce a serious bug that made it past deployment before discovery, it could ruin the day for half of all living adult humans.

For each of the 40 or 50 bugs dealt with back in M128, Google had to validate the bug, fix the bug, QA the fix to make sure it didn't conflict with the billions of web pages and hundreds of thousands of Chrome extensions, release a fix, and then encourage users to update their browsers.

Also: Google Earth added Nano Banana, and I immediately reimagined Philly with zombies and evil clowns

That's a lot to accomplish in roughly a month, even with a developer army that a company like Google can task with the work.

To be sure, Google has been using AI tools to manage these earlier releases. Back in 2023, Google used AIs to increase security fuzzing coverage (basically testing with lots of unexpected, random inputs). By 2024, about the time of M128, Google started to add specialized AI tools to do vulnerability analysis. In 2025, the company said, "We collaborated with DeepMind and Project Zero on Big Sleep, an AI vulnerability discovery agent that successfully found bugs in the V8 JavaScript engine and graphics stack."

Then, in 2026, Google upped its game with an agentic harness (a structure for interacting with LLMs) based on Gemini that was designed to find vulnerabilities across the entire Chrome codebase.

This is not a linear process. It's not just a matter of looking at code and realizing that inputs weren't properly sanitized. Instead, it's much more like a game of chess, except the AI is predicting moves thousands of moves ahead. It's not just looking at the code. It's gaming out scenarios. Billions of scenarios.

Doing this doesn't necessarily find flaws in the code. The code could be rock-solid. Instead, it's finding vulnerabilities, which reflect patterns of interactions that could cause a security failure.

One example of this is a vulnerability called a sandbox escape that has existed in Chrome since 2013. This is a bug that, if exploited, would trick the browser into reading local files. That flaw lived in the Chrome code, passing test suite after test suite, *for more than a decade*. Gemini found it in early 2026.

Also: How to keep your AI conversations as private as possible

But here's the thing. While 40 or 50 vulnerabilities in a month are moderately manageable for a Google-sized team, 1,072 bugs over two releases are nothing short of overwhelming.

Here are the two main implications. First, if the AI can find that many vulnerabilities, perhaps it can help manage, repair, test, and deploy fixes. Second, if Google's AI can find that many vulnerabilities, so can AIs operated by enemy actors. And who's to say that the vulnerabilities Google finds and fixes are the same ones that the bad guys find and exploit?

It's all very disturbing.

An explosive arms race

The thing is, the rate of reaction has to be at warp speed. Billions of people are depending on it. But all those people aren't just counting on a continual flow of fixes; they're counting on those fixes not making things worse.

To manage this, Google has developed a four-stage triage operation.

At Stage 1, AI agents help filter out the bad reports, duplicate reports, and bug reports that don't really describe a Chrome security vulnerability. Google calls this "filtering out the noise." Think of it like going through your email and clearing out the junk.

At Stage 2, AI agents actually reproduce the bugs. Chrome is put through its paces in virtual environments that correspond to the specific browser and operating system reported in the bug. If the bug can be reproduced, the AI adds additional details, like stack traces, to the report.

Also: Open weights vs. closed: An AI civil war's afoot, and the stakes are existential

This is a time-consuming stage when humans have to do it. When I get a bug report for my small product, it can take me hours to try to confirm that what a user is reporting is actually true. That's not fixing time. That's just time to determine if the bug should stay on my to-do list. If Google can get Gemini to do this automagically, the time savings for the coding team can be considerable.

At Stage 3, AI agents "enrich the report with metadata." How many detective shows have we all seen where the detective is handed a police report or folder on the crime? That folder contains all the background information on the case before it is handed off to an investigator. That's what Stage 3 is doing. The AI agents are building that report in preparation for handoff.

Then, at Stage 4, the AI agents identify the right human owner and hand off all the case information to that investigator. This allows the investigator to focus on the investigation and not spend a bunch of time on the administrative details necessary to initiate the case.

Also: I let Chrome's AI agent shop, research, and email for me - here's how it went

Google said, "While it's hard to measure precisely, we estimate that this new process is saving hundreds of hours of developer time per month."

Machines fighting machines. It's really the only way to handle the rapid acceleration in exploit velocity. But the last thing you want is some AI hallucinating a fix that's delivered to half the adults on the planet and then fails.

But still, with thousands of vulnerabilities being found at speed, Google needed to scale up the bug-fixing process. To do this, the blog post describes a multiagent workflow.

Also: Google Search will let you instantly generate AI images for free - here's how

There's a fixing agent that writes code, producing a bunch of what Google calls candidate fixes. Then there's a critic agent. This AI evaluates the fixes, determines which would be the best fit for the problem, and provides additional supporting information for developers to evaluate the fix.

The two agents, the fixer and critic, cycle in a way similar to a traditional code review process until they can be sure the code is functional and meets style guidelines.

Next up is the small test-writing agent army. These agents construct tests and run Chrome through the tests across all of Chrome's supported platforms. In this way, problems can be found and submitted back to the fixer and critic before any human developer has to review the fix.

Google said this process saves weeks of developer time.

But users are stubborn

There is a gap in time between when attackers find and exploit bugs and when the fixes are released. Even when Google knows there's a bug in the wild, it has typically taken weeks for the fix to reach the Chrome stable channel.

Google is working to transition to a two-week delivery cadence for major milestones and a weekly release for security updates. But with the uptick in attack velocity, Chrome is pushing to do two security releases each week.

For almost a decade, Chrome has used a silent background download mechanism that would stage updates on disk, automatically running updates once a user restarted Chrome.

Also: Tired of AI Overviews? I found 9 Google Search alternatives that showed me links again

And yeah, none of us wants to install two new Chrome releases each week. We also don't want to restart Chrome. But that's a problem because the time between when a bug fix is downloaded and when a user restarts Chrome could be long enough for an exploit to take hold.

Google said it has been working on ways to push updates without being intrusive. It is trying out a mechanism for dynamic patching that may eliminate most browser restarts. It is also working on saving the state of the browser locally so it can be resumed easily. Another tactic is to find opportune and nonintrusive opportunities to restart the browser.

Also: Chrome vs. Edge vs. Firefox: I tested each browser's AI, but I'm only sticking with one

As a user, I really don't like it when a vendor decides to nanny me and force updates. But I also don't like it when bad guys try to exploit my machine. It looks like Google is trying to walk that fine line, and it might just work.

Save the browser, save the world

Google is publishing a lot of technical information about how it is integrating AI into its defensive workflows. While the technical details are interesting, it's the strategic implications that are more profound.

Also: Google is training AI on even more of your data now, unless you opt out - here's how

AI has escalated the pace of attack. It can also escalate the pace of defense. The arms race is hotter and more intense than ever before. If Google can identify and find two years of bugs in two months, we can derive the pace of innovation from that ratio.

Code lives in a faster timeline than ever. Two years in two months means that a year of AI-enabled coding, attack, and defense will produce the output that previously took 12 years. That's more than a decade of advancement every year.

It's overwhelming. AI is causing that speed increase, and because of it, we're now dependent on AI to defend us at speed.

*You can follow my day-to-day project updates on social media. Be sure to subscribe to my weekly update newsletter, and follow me on Twitter/X at @DavidGewirtz, on Facebook at Facebook.com/DavidGewirtz, on Instagram at Instagram.com/DavidGewirtz, on Bluesky at @DavidGewirtz.com, and on YouTube at YouTube.com/DavidGewirtzTV.*