Google’s AI agent Big Sleep stops a cyberattack before it starts

For decades, cybersecurity has felt like a hopeless game of catch-up. Attackers find a zero-day, deploy it in the wild, and defenders scramble to push patches after the damage is already done. Google is now trying to flip that script using autonomous AI.

Enter Big Sleep, a specialized AI agent built jointly by Google DeepMind and Project Zero. The watchdog just pulled off an unprecedented save: stopping a live, adversary-ready exploit before a single attack was ever launched.

Beating hackers to the punch

The vulnerability in question, CVE-2025-6965, was lurking inside SQLite, the ubiquitous open-source database engine powering billions of devices worldwide. Google revealed that the flaw was previously known only to threat actors who were actively preparing to weaponize it.

Google's threat intelligence team initially picked up faint indicators of malicious activity but couldn't pinpoint the exact flaw. They handed those limited clues to Google's zero-day team, who set Big Sleep loose on the target. The AI successfully isolated the exact flaw the hackers intended to strike, effectively disarming the weapon before it fired.

While Google hasn't revealed who the threat actors were or what specific telemetry gave them away, this marks a historic milestone. Google believes it is the first time an AI agent has been used to directly foil an in-the-wild cyberattack.

An automated defense roster

Big Sleep isn't an overnight fluke. Back in November last year, the agent bagged its first real-world bug, and it has surfaced several more since, protecting both Google’s internal ecosystem and broader open-source projects.

It is also part of a wider security push inside Google. The company is actively fielding Timesketch, an open-source forensics utility powered by Sec-Gemini, alongside an internal insider-threat detector dubbed FACADE that has run quietly since 2018. Meanwhile, to quell fears of autonomous models going rogue on production code, Google released a white paper detailing how it enforces privacy and keeps its defensive agents from unintended actions.

Governments are taking note too. The U.S. Department of Defense is set to unveil winners of a national contest dedicated to building automated AI shields for vital digital infrastructure.

Why it matters

For developers and system administrators, this is pure upside. Massive shared foundations like SQLite underpin everything from your phone’s mobile apps to enterprise servers. Having an AI agent patch a critical open-source flaw before an exploit hits the wild saves millions of organizations from panic-patching over a weekend.

On the industry level, this shifts the structural economics of cybersecurity. Historically, attackers held the asymmetric advantage: they only needed to find one overlooked crack, while defenders had to secure every inch of surface area. AI agents like Big Sleep balance the scale by scaling human expertise across colossal codebases around the clock.

We think Google’s move to deploy frontier models directly onto open-source defenses is brilliant, but it undeniably kicks off an AI arms race. The exact techniques Big Sleep uses to predict and isolate vulnerabilities will inevitably be adapted by well-funded adversary groups to hunt for 0-days faster than humans can review pull requests. The defense won this round, but the clock is ticking on the next.