Google's AI Agent Harness Found a 13-Year-Old Chrome Flaw — and 1,442 Bugs Across Three Releases

Google's bug-hunting AI reached back thirteen years to prove it works. The agent harness the company's Chrome Security team built in early 2026 — a system running Gemini models across the browser's source code — surfaced a sandbox escape, CVE-2026-3545, that had sat undetected in Chrome's Navigation code for more than 13 years. Google says the flaw could have let a compromised renderer trick the browser into reading local files off a victim's machine. Catching a defect that old is the headline the company wanted. The number underneath it is the one defenders should sit with: 1,442 security fixes across three consecutive releases — Chrome 149, 150 and 151 — more than Google shipped across the previous 23 Chrome milestones combined.

This is where the story we covered when Chrome 151 landed with 370 fixes turns from a single large release into a trend with a mechanism attached. In a July 31 disclosure, Google confirmed what the volume implied: the surge is AI-driven, the tooling now runs across both discovery and patching, and the bugs it is clearing are frequently old. The defender question is no longer which CVE to prioritize. It is whether an organization's Chrome update process can absorb fixes as fast as an AI can generate them.

The Flaw That Validated the System

CVE-2026-3545 is described as an insufficient-data-validation issue in Chrome's Navigation component that enabled a sandbox escape through crafted HTML pages. Both The Hacker News and SecurityWeek tie it to the agent harness and agree it hid for more than 13 years — but they diverge on the particulars, a reminder to treat secondary specifics as reported rather than settled. The Hacker News lists a CVSS score of 9.6 and says Google patched the bug in a March stable-channel update; SecurityWeek reports 9.8 and an early-May fix in Chrome 145. The through-line both confirm is what matters: a critical, decade-plus-old sandbox escape that human review and years of fuzzing had missed, found once an LLM-driven system was pointed at the codebase.

The Number Behind the Bug

Google patched 1,072 defects across Chrome 149 and 150 alone — in its own phrasing, "surpassing the total number of security bugs fixed across the prior 23 milestones combined." Chrome 151 added 370 more, of which 349 were reported by Google itself and seven were rated critical. That brings the three-release haul to 1,442, and the browser's 2026 total past 1,800. The pace is not confined to Chrome: the U.S. National Vulnerability Database has logged 46,872 flaws so far this year, already close to the 49,920 recorded for all of 2025.

Inside the Agent Harness

Google has described the machinery in more detail than most vendors offer. The Chrome team started using large language models in 2023, but the piece that changed the numbers is the agent harness stood up in early 2026. Per the company's disclosure, it was trained on a knowledge base of previously identified CVEs and Chrome's entire Git history, supports swapping models in and out, and reads developer-written SECURITY.md files through a "critic" agent that weighs candidate findings. It can run vulnerability-finding models over the code repeatedly, and Google says it now generates candidate fixes for most of the bugs it surfaces. To limit the risk of the system behaving unexpectedly, Google says the AI analyzes source code strictly at rest, on locked-down machines with no general internet access.

One thing Google has not said: whether the harness, or the models behind it, will be published or open-sourced. The disclosure describes the design without releasing it, so for now the harness should be treated as proprietary and unavailable to defenders who might want to run something similar against their own code (open question).

The Patch-Absorption Problem

What follows is The CyberSignal's assessment, not new reporting. The reflexive reading of a 1,442-fix run is that Chrome is riddled with holes. The more useful reading is close to the opposite: a pipeline clearing hundreds of latent defects at once is defense working at a speed it never could before. The problem the record pace creates sits on the receiving end. AI can now find and draft fixes for decade-old bugs faster than most enterprises can test, approve and deploy a browser update. When discovery was the bottleneck, a slow patch ring was tolerable, because new critical bugs arrived at a human rate. Remove that bottleneck and the exposure shifts almost entirely to the gap between a fix existing and a fix installed.

A 13-year-old flaw is reassuring once it is patched on your fleet and a liability every day it is not — and the supply of such flaws is now machine-paced. It is the same pattern The CyberSignal tracked when Anthropic's Mythos surfaced 271 flaws in Firefox 150: long-latent bugs unearthed in bulk, at a rate defenders never had to plan for. Google's own response — piloting twice-weekly security releases and restart-free "dynamic patching" — reads as an admission that the old cadence cannot hold. Organizations that mirror that thinking will be fine. Those still gating Chrome behind slow manual approval will feel the acceleration as constant pressure.

What This Means for Defenders

The takeaway is unglamorous and time-sensitive. Confirm managed Chrome fleets are on 151 or later and that update policies are not pinning endpoints to an older milestone; with seven critical fixes in that release and no reason to wait, it should move through deployment rings quickly. More durably, treat Chrome's release cadence — not any single CVE — as the thing to plan around, and audit whether your update tooling can handle two security releases a week and background restarts before that becomes the default rather than a pilot. Keep watching CISA's Known Exploited Vulnerabilities catalog as well: none of the 1,442 fixes is reported as exploited in the wild — unlike the actively attacked Chrome V8 zero-day we covered earlier this year — but the more AI-found bugs ship publicly alongside their fixes, the faster attackers can reverse those fixes into exploits aimed at whoever has not updated. Speed of adoption is the control that matters.

Open Questions

Several threads stay open. The CVSS score and exact patch date for CVE-2026-3545 differ between outlets (reported, not resolved). Whether the agent harness or its models will be released publicly is unanswered (open). It is unclear whether other Chromium-based browsers inherit specific CVEs through shared code, and — the question that governs the rest — whether AI-found volume holds at this level or falls back once the backlog of old bugs is cleared. Until a few more releases land, a 1,442-fix stretch could be a durable new baseline or a one-time clearance.

Primary Documents

• Google — "Stronger with every update: making Chrome and the web safer in the AI era" (primary disclosure)
• NVD — CVE-2026-3545
• The Hacker News — "Three Recent Chrome Releases Fix 1,442 Flaws, More Than Prior 23 Updates Combined"
• SecurityWeek — "Google AI Uncovers 13-Year-Old Chrome Flaw Amid Record Patching Pace"