Google's Chrome Security team has disclosed that artificial intelligence tools found and helped fix 1,072 security vulnerabilities in just two Chrome releases — more bugs than the team had squashed across the previous 23 milestones combined. The figures, published in a detailed technical blog post on July 30, 2026, offer one of the clearest demonstrations yet that large language models can operate at the heart of a major software security program.
For anyone following the latest AI developments, the implications stretch well beyond a single browser. Google's disclosure arrives amid a broader industry reckoning over how AI tools are reshaping cybersecurity — for both attackers and defenders.
An Agent Harness Built on Gemini
In early 2026, Google built what it calls an "agent harness" that uses its Gemini family of models to scan the sprawling Chrome codebase for vulnerabilities. The system is designed to find real security flaws with higher efficiency and lower false-positive rates than traditional static-analysis tooling.
One discovery stood out. The AI identified a sandbox escape vulnerability that would have allowed a compromised renderer process to trick the browser into reading local files on a user's machine. The bug, according to Google, had "quietly survived in our codebase for more than 13 years."
"For many of us, this moment cemented the potential of AI-powered vulnerability detection," the Chrome Security team wrote.
From Bug Finding to Automated Triage
The same AI systems that hunt for bugs are now triaging them. Historically, reviewing a single incoming security report consumed between five and 30 minutes of a human engineer's time. Google's automated pipeline now handles the bulk of that work: filtering duplicate reports, attempting to reproduce the bug, assigning a severity rating, and routing the issue to the appropriate engineering team. The company estimates this saves hundreds of hours of developer time every month.
Fixing bugs at scale required rethinking the entire repair process. The current architecture runs a multi-agent pipeline in which a fixing agent generates multiple candidate patches, a separate critic agent evaluates which patch is strongest, and a test-writing agent verifies the fix works across all of Chrome's supported platforms before a human reviewer ever sees it.
Google partnered closely with its DeepMind division and the Project Zero research team on the effort, building on tools internally known as BigSleep and CodeMender. These are natively integrated into Chrome's continuous integration system, running every 24 hours across all incoming code changes to catch security bugs before they ship.
"This integration has yielded significant results," Google stated. "In May alone, we blocked over 20 vulnerabilities from reaching production, including a critical S1+ issue."
External Researchers Feel the Squeeze
The surge in internal AI-driven discovery has had a knock-on effect on Google's external Vulnerability Reward Program (VRP). By March 2026, Chrome received more bug reports from external researchers than it had across all of 2025. In response, Google adjusted the VRP to incentivize findings that go beyond what its own AI tools are already catching — a signal that automated vulnerability discovery is beginning to outpace traditional manual bug hunting.
Multiple outlets, including TechCrunch, SecurityWeek, BleepingComputer, and WIRED, confirmed the figures. WIRED reported that Chrome now requires roughly twice-weekly patching cycles, a cadence made feasible only by the AI pipeline's throughput.
Closing the Patch Gap
Finding and fixing bugs quickly only matters if those fixes reach users in time. Google acknowledged the persistent challenge of the so-called "patch gap" — the window between when a fix lands in the open-source codebase and when it reaches the stable channel most users run. During that window, attackers can reverse-engineer the patch and exploit the flaw before users update.
To compress that gap, Chrome is developing dynamic patching technology that can replace vulnerable background processes on the fly without requiring a full browser restart. The team has already shipped a related change in Chrome 150 that automatically restarts the browser on macOS when all windows are closed and an update is pending.
The longer-term structural work focuses on eliminating entire categories of vulnerabilities rather than patching individual instances. Google is expanding MiraclePtr, a technology that neutralizes use-after-free memory bugs, to additional rendering libraries including Skia, ANGLE, and Dawn. The goal is to cover a growing share of the codebase where these memory-safety issues originate.
A Broader Shift in Software Security
Google's disclosure arrives at a moment of intense scrutiny over AI's dual-use nature in cybersecurity. In the same week, WIRED, ABC News, and Politico reported that Anthropic's Claude models independently breached three real-world organizations during security testing, writing and deploying malware in the process. The parallel developments underscore a central tension: the same class of AI capabilities that can autonomously patch a 13-year-old vulnerability can also be directed toward offensive ends.
For Chrome's team, the calculus is clear. The 1,072 bugs fixed in two releases represent real, user-facing security improvements — vulnerabilities that might otherwise have lingered for years. Whether the rest of the software industry can replicate Google's pipeline, and whether adversaries can weaponize the same techniques faster than defenders can deploy them, remains the defining question of the AI security era.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
