Google (NASDAQ:GOOGL) revealed that its AI-driven cybersecurity agent, Big Sleep, has successfully intercepted and neutralized an active security threat, marking a milestone in automated cyber defense.
According to CEO Sundar Pichai, the AI uncovered a critical vulnerability in SQLite—cataloged as CVE-2025-6965—that had previously gone undetected by the broader security community. The flaw, known only to malicious actors, was on the verge of exploitation when Big Sleep identified it.
Google claims this is the first known instance of an AI system directly preventing the real-time exploitation of a software vulnerability.
Jointly developed by Google DeepMind and the Project Zero team, Big Sleep continuously scans for hidden flaws in widely used software systems. Since identifying its first vulnerability in the wild in November 2024, the AI agent has uncovered numerous security issues, surpassing the company’s initial expectations.
Big Sleep is also being integrated into efforts to strengthen key open-source projects, which Google believes will have a cascading effect on improving cybersecurity throughout the broader internet ecosystem.
The company also previewed several upcoming AI-powered security tools expected later this summer. Among them is an upgraded version of the Timesketch digital forensics platform, now enhanced with Sec-Gemini AI features to speed up threat investigation and incident response. A public demonstration is planned for Black Hat USA.
Additionally, Google will present FACADE (Fast and Accurate Contextual Anomaly Detection), an internal AI-powered system used since 2018 to identify insider threats. At DEF CON 33, Google will join forces with Airbus for a Capture the Flag competition showcasing how AI can amplify the effectiveness of cybersecurity experts.
To support wider industry efforts, Google is donating data from its Secure AI Framework to the Coalition for Secure AI. This move aims to accelerate progress in AI agent development, cyber resilience, and software supply chain security.
Google also confirmed that the final round of its two-year AI Cyber Challenge, co-led with DARPA, will take place at DEF CON 33 next month.