A vulnerability hiding in Chrome’s code since 2013 just got found by a machine. Google’s AI-driven vulnerability detection system, built on its Gemini model, identified a high-severity flaw designated CVE-2026-3545. The bug lives inside Chrome’s Navigation component and qualifies as a sandbox escape, meaning an attacker could use it to break out of the browser’s protective isolation layer and reach the underlying operating system. Its CVSS score is 9.8 out of 10.
A 13-year guest that nobody invited
The flaw was patched in Chrome version 145, released in early May 2026, but the code it exploits had reportedly been sitting there since around 2013. The AI agent that found it uses a combination of a historical CVE database and Git commit history to methodically trace how code has evolved over time, running its analysis on restricted machines that process inactive code in isolation.
Google was careful to frame the AI system as complementary rather than a replacement. The team noted that fuzzing, human researchers, and external reports all remained active contributors to Chrome’s security posture in 2026, with external submissions actually rising enough to prompt adjustments to the Chrome Vulnerability Reward Program.
The patching numbers are almost hard to believe
Chrome versions 149 and 150 together fixed 1,072 security bugs. That single pair of releases outpaced the cumulative total from the previous 23 stable Chrome milestones combined. By late July 2026, total security fixes for the year had crossed 1,800, with Chrome 151 alone contributing another 370 patches to that count.
Google has also moved to a pilot program for twice-weekly security releases, compressing the window between a vulnerability being patched internally and that fix reaching users. Google is working toward automated triage and fix generation, meaning the pipeline from discovery to patch could eventually shrink from days to hours.
What this means for the security landscape
The discovery of CVE-2026-3545 also has implications beyond Chrome. The technique Google’s AI agent used, cross-referencing current code against historical CVE patterns and Git history, is applicable to virtually any large open-source codebase. The Linux kernel, major JavaScript runtimes, cryptographic libraries, and browser engines other than Chrome all carry years of accumulated code that has never been subjected to this kind of retrospective AI analysis.
For crypto infrastructure specifically, this matters. Wallets, nodes, and decentralized applications run inside browsers or depend on browser security models for their interface layer. A sandbox escape vulnerability in Chrome is a potential attack surface for anything operating through a browser extension or web interface. A CVSS 9.8 bug that sat undetected for 13 years in that layer is exactly the kind of thing that should prompt a review of how crypto developers think about browser-level security assumptions.