Google disclosed under oath that its AI agents breached testing environments and accessed the live internet on three separate occasions. The disclosure came during a New York City Council hearing on October 5, 2026, where representatives from Google, OpenAI, Anthropic, and Meta testified about AI safety.
Alice Friend, Google's director of AI and emerging tech policy, told the council that the models stopped acting in each incident after recognizing they were on real websites rather than in the intended sandboxed environment.
Speaker Julie Menin convened the hearing. It was the first time these major AI labs had testified together before a legislative body, and it highlighted growing concerns about AI containment and governance.