Google announced Thursday that its AI-powered tools assisted in fixing 1,072 security vulnerabilities across two Chrome browser versions released in June—Chrome 149 and Chrome 150. That figure surpasses the combined total of 1,036 fixes across the 23 stable releases Google shipped over the preceding two years. The scale matters: Chrome is the world's most widely used browser, running on an estimated three billion devices, which means vulnerabilities left unpatched can affect more users than flaws in nearly any other consumer software product.

Among the discoveries was a sandbox escape vulnerability that had lain dormant in Chrome's codebase for more than 13 years. Google said the flaw, if left unpatched, could have enabled a compromised renderer process to deceive the browser into reading local files. Sandboxing is a core browser security boundary designed to confine untrusted web content so that even if an attacker exploits a rendering bug, they cannot reach the operating system or user data—making an escape from that boundary particularly serious.

Large language models have "fundamentally shifted the economics of cybersecurity, transforming vulnerability discovery into an automated, industrial-scale operation," said Doug Turner, Chrome's director of engineering. He noted that models like Gemini allow the Chrome team to "preemptively" address flaws and stay ahead of potential attackers.

The company emphasized that "an increase in bugs found and fixed is not a sign of failure."

A Multi-Year AI Security Buildout

Google has been deploying large language models against its own security infrastructure for several years. The effort began with fuzzing coverage in 2023, followed by the Naptime tooling developed with Project Zero in 2024. In 2025, DeepMind and Project Zero introduced the Big Sleep discovery agent, which identified bugs in Chrome's V8 JavaScript engine and graphics stack.

In early 2026, Google constructed a Gemini agent harness to systematically scan the broader Chrome codebase. That system uncovered the sandbox escape that had persisted undetected since 2013. The 2026 harness incorporated interoperable open-weight and proprietary models, a knowledge base assembled from Chrome's Git history and past CVEs, and a "critic" agent that reviews developer-authored SECURITY.md files to identify where trust boundaries exist.

Automated Triage and Patching

Google is also training AI to validate, categorize, and repair incoming security reports. Previously, triaging a single report required between five and more than thirty minutes of human attention. A hybrid pipeline combining rules and AI models now filters out spam and duplicate submissions, reproduces bugs on the relevant operating system and Chrome version, attaches metadata such as severity and introduction date, and routes each report to an appropriate owner.

Google estimates this pipeline saves hundreds of developer hours each month.

For repairs, a fixing agent proposes multiple candidate patches, while a second critic agent selects the strongest option—iterating in a manner analogous to a code review until the change compiles successfully and complies with Chromium style guidelines. The system also generates tests before an engineer provides final approval.

The models operate on locked-down machines with no general internet access, run behind an allowlisted network interception layer, and are restricted from accessing files outside designated source directories.

Closing the Patch Gap and Reshaping Bug Bounties

Google is actively working to narrow the "patch gap"—the interval between when a fix lands in public source code and when users actually apply it. During that window, attackers can study patches and develop exploits. To address this, Google is trialing two security releases per week and developing dynamic patching capabilities that would allow components of Chrome to be updated without requiring a full browser restart.

The surge in AI-assisted discovery has also reshaped Google's bug bounty program. External submissions grew so rapidly that by March, the company had received more reports than in all of 2025 combined. This prompted a revamp of the Chrome Vulnerability Reward Program to direct researchers toward the categories of bugs least likely to be found by Google's own automated tools—a signal that automation is already shifting the value of human security research toward areas machines still struggle to reach.

Over the longer term, Google is expanding its MiraclePtr memory-safety tool, converting new components to Rust, and donating $12.5 million to the Alpha-Omega open-source security project, an initiative housed under the Linux Foundation that funds security improvements for widely used open-source software.

For comparison, Microsoft's most recent Patch Tuesday addressed a record 570 flaws and referenced AI in its process, while Apple has independently recorded 482 fixes in 2026—though neither rival appears to be on the same trajectory as Google's AI-driven approach. The gap suggests that companies with deep AI research capabilities are increasingly able to leverage those assets for defensive security at a pace competitors without equivalent in-house models may find difficult to match.