Google said artificial-intelligence tools have sharply increased the number of security flaws its Chrome team can find, evaluate and repair, with the browser's two most recent stable milestones addressing more vulnerabilities than the previous 23 milestones combined.
Chrome versions 149 and 150 fixed 1,072 security bugs, according to the company. Google attributed the increase to using large language models at several stages of the vulnerability lifecycle, from source-code analysis to triage and proposed patches. The company estimates its automated triage work is saving developers hundreds of hours each month.
The effort developed over several years. Google said it used language models to expand fuzzing coverage in 2023, worked with Project Zero on the Naptime vulnerability-research system in 2024, and collaborated with DeepMind and Project Zero on the Big Sleep agent in 2025. In early 2026, the Chrome team built a Gemini-based agent harness to examine the broader codebase with fewer false positives.
One result was the discovery of a sandbox-escape flaw that had remained in Chrome for more than 13 years. The bug could have allowed a compromised renderer to induce the browser to read local files, Google said. The company's systems now generate candidate repairs for most vulnerabilities, while Big Sleep and CodeMender run daily within continuous integration. In May, those integrated checks prevented more than 20 flaws, including one rated critical, from reaching production.
Google stressed that the AI analysis operates in restricted environments. Models examine source code on locked-down computers without general internet access. Network requests are intercepted and controlled through application-and-destination allowlists, and agents are constrained from modifying the wider system or reading beyond designated source directories. Conventional fuzzing and external researchers remain part of the program, particularly for complex flaws involving interactions across distant parts of the code.
Finding more bugs also creates pressure on release operations. Once a repair becomes visible in an open-source repository, attackers can study it before the change reaches most users, creating a patch gap. Chrome is moving toward major milestones every two weeks, with weekly security updates, and Google said it was piloting two security releases per week. It is also automating release notes and descriptions for Common Vulnerabilities and Exposures while retaining public disclosure for flaws that reach Stable.
The figures are Google's own account of its security program rather than an independent audit, and a larger number of fixes can reflect greater detection as well as underlying defect volume. Still, the scale of the reported increase illustrates a practical consequence of AI-assisted coding: defensive teams are applying the same automation to discovery and remediation, while trying to preserve containment, human oversight and timely delivery.
