Google has announced that its AI-assisted security workflows have helped Chrome fix 10,721 security vulnerabilities across Chrome Stable milestones 149 and 150. This number exceeds the total number of bugs patched across the previous 23 milestones combined.
In a new report, the Chrome Security Team detailed how large language models are integrated throughout the vulnerability lifecycle, from discovery and triage to patch development, testing, release, and update adoption.
This initiative reflects Google’s commitment to countering an evolving threat landscape, where attackers also leverage AI to accelerate vulnerability research and exploit development.
Google Uses AI Fix 1,072 Chrome Flaws
Chrome’s AI-powered vulnerability discovery capabilities build on earlier projects, including AI-assisted fuzzing, Project Zero’s Naptime framework, and Big Sleep, a collaboration between Google DeepMind and Project Zero.
In early 2026, Google developed an agent utilizing Gemini to scan the broader Chromium codebase more efficiently and with fewer false positives.
One significant finding was a sandbox escape vulnerability that had reportedly been present in Chrome’s code for over 13 years.
This flaw could have allowed a compromised renderer process to convince the browser to read local files, showcasing AI’s potential to identify long-standing security defects in complex codebases.
According to Google, its updated system uses both proprietary and open-weight models, supplemented by a knowledge base that includes Chromium’s Git history and previously reported CVEs.
The company has also encouraged developers to add SECURITY.md files, which document trust boundaries and threat models, providing AI agents with additional context to identify legitimate security risks.
To mitigate operational risks, Google runs these models in restricted environments. The systems analyze source code on locked-down machines without unrestricted internet access, with network activity intercepted and controlled through strict allowlists.
Beyond discovery, Google is automating the security bug triage process through a pipeline that filters out spam and duplicates, validates proof-of-concept exploits, reproduces flaws on affected configurations, assigns severity ratings, and routes reports to the relevant component owners. Google estimates that this automated workflow saves developers hundreds of hours each month.
For remediation, multi-agent workflows generate candidate patches, review proposed changes, and produce tests before human developer approval.
Fixing and critic agents simulate elements of a code review process, ensuring that the code complies with Chromium standards and local conventions. Test-writing agents also validate patches across Chrome’s supported platforms and configurations.
Google stated that its AI tools are integrated into continuous integration systems and scan code changes every 24 hours. In May alone, the company blocked more than 20 vulnerabilities from reaching production, including one critical severity issue.
The Chrome team is also addressing the “patch gap”, the time between a public code fix and its widespread deployment to user devices. Google is moving towards major Chrome releases every two weeks, weekly security updates, and is piloting two security releases per week.
Additionally, it is researching dynamic patching, which would enable updates to renderer and GPU child processes without requiring a full browser restart.
The company’s long-term strategy includes expanding C++ memory-safety mitigations, adopting Rust for high-risk components, automating third-party dependency updates, and utilizing AI in Chrome’s commit queue to detect vulnerabilities before they enter the codebase.
ALERT: 20+ government sites delivered malware to businesses and citizens. See full attack research to check your own exposure.
