Google says it has developed an AI-powered cybersecurity tool that can stop an attack before it is launched

Google

Google is unveiling a new approach to cybersecurity — one where artificial intelligence takes centre stage not just as a tool, but as the first line of digital defence.

In the lead-up to the cybersecurity industry’s biggest summer events, Google has offered a glimpse into its fast-growing suite of AI-driven security solutions, signalling a clear shift in strategy: from reactive measures to proactive, intelligent threat detection and mitigation.

In a blog post, Google noted that one of the key tools is is Big Sleep, an AI agent developed by Google DeepMind in collaboration with Project Zero. Designed to autonomously detect software vulnerabilities, Big Sleep made headlines last year when it uncovered its first real-world exploit. Now, it’s being credited with something unprecedented: stopping a security threat before it was launched.

Using intelligence from Google’s Threat Intelligence division, Big Sleep identified a critical SQLite flaw (CVE-2025-6965) that was known only to attackers. Google says this marks the first time an AI agent has actively foiled an exploit in the wild. The tool is also being applied to strengthen the security of widely-used open-source software, vastly increasing the scale and speed of defensive coverage.

To ensure safety and accountability, Google has released a white paper detailing its secure-by-design approach to building AI agents focusing on human oversight, transparency, and privacy.

In addition to Big Sleep, Google is infusing AI into its security infrastructure:

•Timesketch, the company’s open-source forensics platform, now integrates Sec-Gemini-powered agents to automate incident response.

•FACADE, an insider threat detection tool, processes billions of security events each day using contrastive learning, without relying on historical attack data.

•A new AI-assisted Capture the Flag challenge at DEF CON 33 will give participants hands-on experience working alongside AI agents in real-time cyber defence scenarios.

Beyond internal development, Google is doubling down on collaboration. It has committed data from its Secure AI Framework (SAIF) to the Coalition for Secure AI (CoSAI) and is preparing to conclude its two-year DARPA-backed AI Cyber Challenge at DEF CON 33 — a competition aimed at creating AI tools to detect and patch vulnerabilities in critical open-source software.