Google's Chrome Security team has reported a significant acceleration in vulnerability detection and patching, attributing the improvement to the integration of artificial intelligence models into their development pipeline. In the last two Chrome releases alone, 1,072 security bugs were fixed, a number exceeding the total fixes across the preceding 23 milestones combined.
The company stated that in early 2026, an agent harness utilizing the Gemini AI model was deployed to identify vulnerabilities across the Chrome codebase. This system demonstrated higher efficiency and reduced false positives compared to previous methods. One notable discovery was a sandbox escape vulnerability that had persisted in the codebase for over 13 years, allowing a compromised renderer to trick the browser into reading local files.
Beyond detection, AI tools are also being used for bug triaging. Historically, human security analysts spent between 5 and 30 minutes triaging a single report. The automated system now handles tasks such as filtering duplicates, reproducing bugs, assigning severity levels, and routing issues to the appropriate teams. This automation is estimated to save hundreds of hours of developer time monthly.
The bug repair pipeline has also been re-engineered. A fixing agent generates multiple candidate patches, which are then evaluated by a separate critic agent. A test-writing agent verifies the effectiveness of the chosen fix across all Chrome platforms before human review. These tools, developed in partnership with Google DeepMind and Project Zero, including BigSleep and CodeMender, are integrated into Google's continuous integration (CI) system and run daily across all change lists to proactively detect security issues. In May alone, this system prevented over 20 vulnerabilities, including a critical S1+ issue, from reaching production.
The increased efficiency has also impacted Google's external vulnerability reward program (VRP). By March 2026, Chrome received more bug reports than in all of 2025, prompting a revision of the VRP to encourage researchers to focus on vulnerabilities not already identified by internal AI tools.
Google is also addressing the "patch gap," the period between when a fix is committed to the open-source codebase and when it reaches users. To minimize this window, which can take weeks for the Chrome Stable channel, the company is developing dynamic patching that allows background processes to update without requiring a full browser restart. Chrome 150 on macOS has already implemented a change that automatically restarts the browser when all windows are closed and an update is pending.
Long-term structural improvements are also underway to eliminate entire categories of bugs. MiraclePtr, a technology designed to neutralize use-after-free vulnerabilities, is being expanded to more libraries, including Skia, ANGLE, and Dawn, with a goal of covering up to 90% of such bugs on the GPU main thread. A separate "spanification" effort has ensured that 97% of first-party Chrome code compiles cleanly with strict unsafe-buffer warnings, moving spatial safety enforcement to the compiler rather than relying on runtime checks.
Acknowledging the limitations of runtime mitigations, Google is also exploring rewriting high-risk components in Rust. A centralized Rust SDK is being developed to facilitate this transition. The company also highlighted the challenge of managing over 2,300 third-party dependencies across Chromium and satellite projects, with approximately 1,700 shipping to users. These dependencies are being moved to automated update pipelines that continuously roll them to their latest upstream versions, with safety signals from Google's Open Source Security Intelligence Platform acting as guardrails.
