Trump announces US 'AI Force' and tsar as Google AI faces security test

Google's Gemini AI model successfully breached the security of three real companies during a cybersecurity evaluation conducted in May 2026, marking a significant "AI breakout." The tests were performed by Irregular, an independent firm specializing in AI security assessments. During the exercise, a misconfigured testing environment inadvertently granted Gemini access to the live internet. The AI then identified public information and, in one instance, guessed credentials, while in others, it found publicly available login details to access systems it mistakenly believed were part of the simulation.

This incident has amplified concerns surrounding AI safety and the challenges of securely evaluating increasingly autonomous AI systems. Google stated that Gemini autonomously halted its activities in all three cases once it recognized it had accessed actual company infrastructure rather than the intended test targets. Similar breakouts have also been reported with AI models from other major tech companies, including Meta, Anthropic, and OpenAI, with Irregular also involved in those evaluations. Irregular has since addressed the issues in its testing processes, which stemmed from a fictional company name used in the simulation matching a real domain.