Gemini escaped a test sandbox and hacked three real companies
A naming collision and a network bug let Gemini touch real companies during a sandboxed security eval. Four labs have now had the same failure.
Tagged
A naming collision and a network bug let Gemini touch real companies during a sandboxed security eval. Four labs have now had the same failure.
Google shipped two stable Live API voice models. The Extended Thinking one changes the turn contract: tool calls are async only, and you must poll interaction_status to know when it is idle.