AI helps Chrome fix 1,072 security bugs

Google says AI helps Chrome to spot and address security issues faster. In the two most recent Chrome releases, the company reports that 1,072 security bugs were patched in total. That number is higher than the combined total solved across the earlier 23 Chrome milestones.

The core of that acceleration lies in a broader use of large language models (LLM’s) across the vulnerability management process: from discovery to generating tests and possible patches.

From searching to patching: AI across the full vulnerability chain

Where security teams often look for automation support at one or two points, Google says it’s now using LLM’s in multiple steps. According to the company, models are not only involved when analyzing reports, but also when drafting and improving the reporting itself.

Google highlights, among other things:

• discovering potential vulnerabilities
• reproducing findings so the bug can be demonstrated
• determining severity and assigning bugs to developers
• generating candidate patches
• creating tests to verify the fix

One important detail: Google says these multi-agent AI workflows help existing security testing, rather than fully replacing it. According to Google, fuzzing remains effective for uncovering complex vulnerabilities.

AI projects behind the scenes: fuzzing, Naptime, and Big Sleep

Google started earlier with AI for security-related tasks. In 2023, the focus was on improving security fuzzing. Later, Google worked with Project Zero and developed the Naptime system, giving AI models access to specialized tools for vulnerability research.

After that came collaboration with, among others, Google DeepMind and Project Zero around Big Sleep. This agent system is said to have found vulnerabilities in parts of Chrome, including the V8 JavaScript engine and parts of graphics.

In early 2026, Google also built a Gemini-powered agent aimed at searching a broader portion of the Chrome codebase. The system particularly wants to reduce false positives, so reports better match what teams can truly investigate.

Example: a sandbox escape that lingered for years

Google cites, as an illustration, a vulnerability that was surfaced by one of the AI systems. It involved a Chrome sandbox escape that is said to have been present in the codebase for more than 13 years.

If an attacker had figured out how to exploit it, a compromised renderer could escape the sandbox and mislead the browser process so the browser could read local files.

Although details like these often become fully clear only after researchers and engineers confirm the bug, the example underscores the broader point: AI can help find vulnerabilities that don’t stand out automatically during traditional review cycles.

Faster triage and reporting: less noise, more speed

Google also sees an increase in reports coming through the Chrome Vulnerability Reward Program. In March 2026, there were even expected to be more security reports submitted than throughout all of 2025.

To keep that flow manageable, Google reportedly adjusted the program. The focus is on reports that “build on” what Google is already finding and processing via automated tooling.

In addition, Google is automating parts of the triage, including:

• filtering spam and duplicates
• reproducing proof-of-concept exploits
• assigning severity ratings
• routing reports to the right developers

Google estimates that this approach saves developers hundreds of hours every month.

Multiple patches, extra review, and less time-to-fix

Once a vulnerability is confirmed, the systems generate multiple possible solutions, Google says. Then another agent reviews the proposed fixes and provides additional information for developers—so they can decide more quickly what works best.

Google reports that this approach has, in May, allegedly prevented more than 20 vulnerabilities from reaching production, including an issue classified as critical.

Not just finding: patches must reach users faster

The company emphasizes that faster patching isn’t only about better detection. Once a fix is in the public source code, attackers can inspect the change and potentially reverse-engineer parts before the security update truly reaches users.

To reduce that “patch gap,” Google is moving to a two-weekly major release cycle with weekly security updates. In addition, the message says pilots are underway for two security releases per week.

Google is also working on dynamic patching: the idea that Chrome can apply updates without having to restart the full process. Starting with Chrome 150 on macOS, Chrome can automatically apply a pending update by restarting the program when the browser is running in the background without open windows.

In the long run, Google wants Chrome to stay continuously up to date with dynamic patching, automatic restarts during inactivity, and improved session restoration.

SECURITY.md and AI-driven security thresholds

Google encourages developers to add SECURITY.md files. These can describe which trust boundaries are relevant and what the threat model looks like. According to Google, that helps AI systems better identify which actions have security implications.

This fits the idea that AI should not only “search,” but also better understand where in the code and workflows it really gets security-critical.

Conclusion: AI accelerates Chrome’s security, but delivery remains crucial

With the approach Google describes, AI helps Chrome in a tangible way: LLM’s support finding, reproducing, prioritizing, and fixing vulnerabilities, including generating tests and candidate patches. At the same time, the biggest challenge seems to shift: it’s not only about finding bugs faster—it’s about getting the solution to users faster and more safely.

By accelerating releases and investing in dynamic patching, Google is trying to shrink the time between “fix on paper” and “fix in the real world.”