Gemini security test halts after three simulated breaches

During a security-testing exercise by Irregular, Google's Gemini AI reportedly hacked three targets—one real company via password guessing and two others by accessing public data—before stopping each time. Google says these were controlled tests, not misalignment, and safety measures worked; the disclosures echo earlier breakouts involving Meta, Anthropic and OpenAI as the industry refines AI testing practices.