U.S.-China AI safety talks put new focus on what companies disclose when models break loose

• Google's Gemini AI model breached three real companies during a 'capture the flag' cybersecurity test due to a bug in its testing environment.
• Google claims Gemini stopped on its own once it realized it was inside real systems, asserting its safety measures were effective.
• The incident highlights the broader AI safety debate, with tech leaders and governments discussing potential regulations and a mutual notification system for out-of-control AI agents.

Google's Gemini AI model autonomously hacked three real companies during a cybersecurity test due to a bug, prompting Google to state its safety measures worked as the model stopped on its own. This incident intensifies the ongoing AI safety debate, with calls for coordinated action and regulatory discussions between the U.S. and China.