Google's Gemini-Based AI Agents Found and Fixed 1,072 Chrome Security Bugs in Two Release Cycles

Bug fixes jump 20x in two release cycles

Google's Chrome security team fixed 1,072 vulnerabilities across milestones M149 and M150, according to a blog post published by the team last week. That figure exceeds the combined total of bug fixes across the prior 23 milestones, Google said.

The growth trajectory is steep. Chrome milestone M128, released Aug. 20, 2024, addressed roughly 50 security bugs — a pace that held through M145. M146 jumped to about 80 fixes. M147 hit roughly 130. M148 reached approximately 350. Then M149 and M150 combined for 1,072.

Chrome holds 73% of global browser market share, with an estimated 3.5 billion active users, according to the report.

An agentic harness built on Gemini

Google's AI-assisted security workflow has been building since 2023, when the company said it used AI to expand fuzzing coverage — automated testing with randomized, unexpected inputs. By 2024, around the time of M128, Google added AI tools specifically for vulnerability analysis.

In 2025, Google worked with DeepMind and Project Zero on Big Sleep, an AI vulnerability-discovery agent that the company says found bugs in the V8 JavaScript engine and Chrome's graphics stack. In 2026, Google deployed what it describes as an agentic harness built on Gemini, designed to scan vulnerabilities across the entire Chrome codebase rather than isolated components.

According to Google, the system found a sandbox-escape vulnerability that had existed in Chrome's code since 2013 — a flaw that could trick the browser into reading local files and had passed test suites for more than a decade before Gemini flagged it in early 2026.

Four-stage triage pipeline

Google described a four-stage AI triage process now in use:

1. Filtering — AI agents screen out duplicate, invalid, or non-security bug reports.
2. Reproduction — Agents recreate reported bugs in virtual environments matching the specific browser version and OS, adding stack traces when reproduction succeeds.
3. Enrichment — Agents attach metadata to build a complete case file.
4. Routing — Agents identify the appropriate human engineer and hand off the enriched report.

Google claims this process saves "hundreds of hours of developer time per month," though the company notes the figure is difficult to measure precisely.

What this means

The jump from roughly 50 bugs per release to 1,072 across two milestones is not necessarily evidence that Chrome got less secure — it likely reflects the AI system's ability to surface latent vulnerabilities that traditional testing and fuzzing missed for years, including the 2013-era sandbox escape. That's a meaningful capability: finding bugs before attackers do is the entire point of a security program.

But the scale creates a new operational problem. A team that could manage 50 fixes per month through manual QA and staged rollout now faces more than 20 times that volume. Google's four-stage triage pipeline is essentially an admission that human review capacity, not bug detection, is now the bottleneck. Whether the QA and deployment process can keep pace with AI-driven discovery — without introducing regressions into a browser used by roughly half the world's adult population — is the harder question the blog post doesn't fully answer.

There's also a dual-use problem Google can't sidestep: if a Gemini-based agent can find a decade-old sandbox escape by reasoning through code paths at scale, comparable tooling in the hands of an adversary could find the same class of bugs first. Google's public disclosure of this capability functions simultaneously as a defensive milestone and a signal to attackers that automated vulnerability discovery at this scale is now feasible.