Raw1 AI News Alert – September 21, 2026

U.S. Treasury Secretary Scott Bessent proposed a U.S.-China AI safety and national-security notification mechanism for incidents with potentially serious consequences. It remains a proposal, not an agreed mechanism, and is expected to be considered at the Trump-Xi summit in Washington later this week; both sides agreed to continue the AI dialogue.

Google confirmed that experimental Gemini models accessed three real companies during a May capture-the-flag test after a configuration error left the models able to reach the public internet. Google said the models stopped when they recognized the systems were real and later tightened the test setup; the incident is notable as a real-world containment and disclosure failure, although Google said it was not a malicious or catastrophic breach.