AI Helped Chrome Close 1,072 Vulnerabilities Across Two Releases

Google has revealed that artificial intelligence has substantially increased the number of vulnerabilities the company is able to find and fix in Chrome. According to its figures, Chrome 149 and 150 closed 1,072 security bugs — more than the browser's previous 23 releases combined.

Today, Google says, large language models are involved at every stage of dealing with vulnerabilities: detection, reproducing reports, assessing severity, distributing tasks among developers, drafting patches and writing tests.

From Naptime to Big Sleep

The journey began in 2023, when LLMs were hooked up to fuzzing. Then, together with the Project Zero team, came Naptime, a system that gave the models specialized tools for vulnerability research. Later, Google DeepMind and Project Zero built the Big Sleep agent, which found defects in the V8 engine and graphics components. In early 2026 the company assembled an agentic harness built on Gemini that combs through all of Chrome's code and cuts down on false positives.

One of the finds is striking: a sandbox escape that had quietly lived in the code for more than 13 years. If exploited, it would have let a compromised renderer break out of isolation and force the browser to read local files.

Google stresses that AI complements rather than replaces the earlier methods — that same fuzzing is still good at catching complex bugs. Developers are asked to add SECURITY.md files describing trust boundaries and threat models, so that the algorithms have an easier time recognizing sensitive operations.

A Flood of Reports and Relief for People

The influx of submissions to the bug bounty program has also surged sharply: by March 2026 Google had received more reports than in all of 2025. The company had to retune the program to prioritize finds that complement its own automated tools. Triage is being automated too — filtering out spam and duplicates, reproducing PoCs, assessing severity, routing. By Google's estimate, this saves developers hundreds of hours a month. In May, the agents reportedly kept more than 20 vulnerabilities out of production, including one critical one.

A Race with the Attackers

Finding bugs quickly gives rise to a new problem: as soon as a fix lands in the open-source code, attackers can study the change and reconstruct the vulnerability before the update reaches users. To narrow this gap, Google is moving Chrome to a two-week cycle for major releases with weekly security updates, and is testing two such releases a week.

To make updates less disruptive, “dynamic patching” is being developed — applying fixes without restarting the browser. Starting with Chrome 150 on macOS, the browser can restart itself to apply an update if it is running in the background with no windows open. The long-term goal is to keep Chrome continuously up to date through dynamic patches, automatic restarts during idle periods and improved session restoration.

Source: Bleeping Computer