Artificial intelligence has moved beyond being a promise and has become the driving force behind security in the world's most popular browser. Google announced that, during the month of June, Chrome patched 1.072 security vulnerabilities in versions 149 and 150 , a figure that surpasses the total number of patches applied in the previous 23 versions combined. This milestone is no coincidence: the company attributes the leap to an AI agent system based on Gemini that has transformed how vulnerabilities are found, classified, and repaired.
The Chrome security team has been using machine learning for years, but the arrival of large-scale language models (LLM) has allowed them to automate processes that previously required hundreds of hours of manual work . According to Google's own data, in May alone they prevented more than 20 vulnerabilities from reaching production, including a critical flaw that was already being exploited . This news has been reported by multiple media outlets, which agree that AI is redefining cybersecurity on an industrial scale.
An unprecedented record in corrections
The numbers speak for themselves. Chrome versions 149 and 150, released in June, include 1.072 security patches, while the previous 23 versions combined totaled 1.036. This means that more vulnerabilities have been patched in a single month than in the previous two years . The technical report published by Google details that Large-Scale Vulnerability Managers (LLMs) are unlocking unprecedented capabilities for automated vulnerability discovery, pushing the boundaries of human expertise. The company emphasizes that this approach does not replace traditional methods, but rather complements them, enabling much broader source code coverage.
The Vulnerability Rewards Program (VRP) has also felt the impact. By March 2026, Google had already received more reports than during all of 2025, forcing them to modify the program's terms to prioritize findings that offer something different from what automated tools already detect. AI filters spam, replicates proofs of concept, and assigns severity levels , saving developers hundreds of hours of work each month.
The AI ​​agent that found a 13-year-old flaw
One of the most striking findings of the new system was the discovery of a sandbox escape that had been hidden in Chrome's code for over 13 years. Cataloged as CVE-2026-3545, with a CVSS score of 9.8 (critical), this flaw would have allowed a compromised renderer to trick the browser into reading local files from the computer. The Gemini-based agent, created in early 2026, was responsible for detecting it by analyzing Chrome's complete source code at a level of detail that human teams cannot sustainably maintain.
Google explains that the system combines multiple specialized agents: some scan the codebase looking for semantic patterns, others validate that the vulnerability is real, and still others generate candidate patches. Human review remains mandatory before applying any changes , but AI greatly accelerates the process. Doug Turner, director of engineering at Chrome, stated that LLMs have fundamentally changed the economics of cybersecurity, transforming vulnerability discovery into an automated, industrial-scale operation.
How the Gemini agent system works
Google's system isn't a single model, but rather an orchestration of agents working in a chain. Detection agents scan Chrome's code for patterns that might indicate vulnerabilities, using semantic code understanding, not just text patterns. Validation agents then confirm whether the finding is real or a false positive. Once confirmed, fixing agents generate multiple potential patches , while patch evaluation agents compare them against the Chromium project's style rules and conventions.
This entire process runs on code stored on locked machines, without general internet access, to mitigate risks. Google has also integrated AI tools into its continuous integration (CI) system, which runs every 24 hours at all build levels. This allows for the proactive detection of security flaws before they reach production . The company states that the models already generate remediation proposals for most of the vulnerabilities processed, supported by the fact that Chrome installs a 4GB AI model to optimize local tasks.
Automating the patching process
AI not only finds vulnerabilities, but also participates in the entire vulnerability lifecycle. From detection to patch generation, including report classification and task assignment, the system filters spam and duplicates, replicates proof-of-concept tests, assigns severity levels, and routes each report to the responsible team . Google estimates that this process saves its developers hundreds of hours of work each month.
Once a patch is ready, it's integrated directly from the main repository into the active stable branch of Chrome, which is continuously monitored to prevent new vulnerabilities. The company has also begun automating the generation of release notes and CVE descriptions , eliminating manual bottlenecks. In May, this system prevented more than 20 vulnerabilities from reaching production, including one classified as critical.
Faster updates and dynamic patching
Finding and fixing a vulnerability doesn't immediately protect users. Once the modification appears in the open source code, an attacker can study it before the patch reaches all computers. To reduce this window of exposure, Google Chrome is changing its pace and updating every two weeks with weekly security updates, and is already testing releasing two security updates per week. Furthermore, it's working on a dynamic patching system that would allow updates to be applied without needing to restart the browser.
Starting with Chrome 150, the browser can now automatically restart on macOS to install a pending update when running in the background with no windows open. The long-term goal is to keep Chrome continuously updated, combining dynamic patching, automatic restarts during periods of inactivity, and improved session restoration that prevents losing open tabs. Google is also transitioning to memory-safe languages ​​like Rust to eliminate entire classes of security vulnerabilities.
The company has confirmed that the Chrome security team began using LLM in 2023, collaborated with DeepMind and Project Zero on tools like Big Sleep, and in early 2026 created the Gemini-based agent that has revolutionized vulnerability detection. This trend isn't unique to Google: Microsoft also patched a record 570 vulnerabilities in June thanks to AI , while Apple maintains similar figures to previous years, suggesting that integrating AI into security makes a significant difference.
The rate of discovery has been so high that the team has begun releasing security patches twice a week, a higher frequency than is typical in the industry. Google estimates that the automated classification saves hundreds of hours of work per month , although it acknowledges that this measurement is not precise. The company emphasizes that humans remain involved: AI-generated patches are submitted to developers for review and final approval, and the system does not automatically merge patches.
The discovery of the 13-year-old flaw demonstrates AI's potential to find problems that traditional methods miss. Google plans to continue using AI to prevent new errors by identifying and eliminating them as close as possible to the time of code submission. The company is also neutralizing issues in the vulnerability tree that could impact code considered secure in isolation, a strategy to prevent chaining vulnerabilities.
For users, the practical recommendation remains the same: Chrome downloads its updates automatically, but still needs to restart to apply most of them. It's advisable to heed the update notification and close and reopen the browser when appropriate. Chrome's security has taken a significant leap forward thanks to AI , but constant vigilance is still necessary in an ecosystem where attackers can also use the same technology to find vulnerabilities.
