Google AI finds 1,072 Chrome bugs in 60 days

Google says its AI security workflow helped find and fix 1,072 Chrome vulnerabilities across two releases, more than the total number of security bugs fixed during the previous 23 milestones combined. The milestone releases, labeled M149 and M150, were completed in roughly 60 days.

The scale matters. Chrome and its open-source Chromium project account for about 73% of the global browser market, with approximately 3.5 billion active users. Google’s Chrome security team detailed the work in a blog post describing how AI is being used to improve vulnerability discovery, triage, and patching.

Chrome’s security workload takes a sharp turn upward

Google’s milestones generally arrive about once a month. Starting with M128, released on August 20, 2024, the company fixed roughly 50 security bugs per milestone through M145. The numbers then rose rapidly: M146 fixed about 80 bugs, M147 about 130, and M148 about 350.

M149 and M150 together reached 1,072 fixes, producing what Google described as a hockey-stick growth curve. The increase presents an operational problem as well as a security opportunity: every vulnerability must be validated, repaired, tested against Chrome’s supported environments, released, and delivered to users.

Recommended reading

PyPI package 999.9.9 harvested SSH keys

A flawed fix could create new problems across billions of browser installations, hundreds of thousands of Chrome extensions, and the web’s vast collection of pages. Google therefore has to increase the speed of its response without lowering the quality of its review.

From fuzzing to Gemini vulnerability agents

Google began using AI to increase security fuzzing coverage in 2023. Fuzzing feeds software unexpected or random inputs to expose failures. Around the time of M128 in 2024, the company added specialized AI tools for vulnerability analysis.

In 2025, Google said it worked with DeepMind and Project Zero on Big Sleep, an AI vulnerability-discovery agent that found bugs in Chrome’s V8 JavaScript engine and graphics stack. In 2026, Google introduced an agentic harness based on Gemini, designed to search for vulnerabilities across the entire Chrome codebase.

The system is intended to reason through chains of interactions rather than simply flagging a line of code with inadequate input handling. Google described it as exploring scenarios that could lead to a security failure, including a sandbox escape that had existed in Chrome since 2013. If exploited, that flaw could trick the browser into reading local files. Gemini found it in early 2026 after it had survived test suites for more than a decade.

The source does not provide a breakdown of how many of the 1,072 vulnerabilities came from each AI system, nor does it identify the full set of tools used in the two releases.

Four stages of AI-assisted triage

Google has built a four-stage process to handle the increased volume of reports:

• Filtering: AI agents remove invalid, duplicate, or irrelevant reports—the “noise” in the intake process.
• Reproduction: Agents run Chrome in virtual environments matching the reported browser and operating system to determine whether the bug can be reproduced. When successful, they add details such as stack traces.
• Enrichment: Agents attach metadata and other background information to prepare a fuller case file.
• Handoff: The system identifies the appropriate human owner and passes along the assembled information.

“While it’s hard to measure precisely, we estimate that this new process is saving hundreds of hours of developer time per month.”

Google says the workflow saves developer time by shifting administrative and verification work to agents. Developers can then focus on investigating and repairing confirmed vulnerabilities instead of determining whether each report belongs on the team’s to-do list.

Fixer, critic, and testing agents

The repair workflow uses multiple agents rather than asking one system to write and submit a patch. A fixing agent produces candidate code changes. A critic agent evaluates those candidates, selects the best fit, and supplies additional information for human developers.

The fixer and critic repeat this process in a cycle resembling code review. They check whether the proposed change is functional and follows Chrome’s coding standards. Separate test-writing agents then construct tests and run Chrome across its supported platforms. Problems found during testing are returned to the fixer and critic before a human reviews the patch.

Google said this multiagent process saves weeks of developer time, although the source does not specify how many fixes were ultimately accepted without human code review or how many candidates each vulnerability generated.

Chrome is targeting twice-weekly security releases

The speed of discovery creates a second bottleneck: delivering fixes to users. Google said that, even when it knows a vulnerability is being exploited, fixes have typically taken weeks to reach Chrome’s stable channel.

The company is working toward a two-week cadence for major milestones and a weekly cadence for security updates. Because attack velocity has increased, Chrome is also pushing toward two security releases each week.

Chrome has used silent background downloads for almost a decade. Updates can be staged on disk and applied when the user restarts the browser, but that creates a window between downloading a fix and activating it. Google is testing several ways to reduce that delay:

• Dynamic patching that could eliminate most browser restarts.
• Saving browser state locally so sessions can be resumed easily.
• Finding less intrusive opportunities to restart Chrome.

Google did not state when the twice-weekly security-release target will become standard. The central tension remains unchanged: Chrome must patch faster as AI accelerates vulnerability discovery, while ensuring that the fixes themselves do not introduce failures at global scale.