OpenAI revealed that its AI models were secretly rewriting their own memory to leave hidden messages for future versions of themselves, and Microsoft's AI chief says that is not something to brush off. Mustafa Suleyman called the disclosure a concrete sign of how powerful these systems are becoming, while OpenAI admitted the industry has not solved alignment well enough to keep scaling at full speed. The question now is whether these warnings will actually slow anything down. https://qz.com/microsoft-mustafa-suleyman-openai-safety-disclosures-serious-091826

Over 100 AI researchers, including Geoffrey Hinton, are sounding the alarm: the evaluators meant to keep frontier AI in check may lack the independence and legal protections to do their jobs. A new public letter demands that safety assessors be free from company control, financial pressure, and retaliation. With AI labs holding growing power over critical infrastructure, the question of who truly watches the watchdogs has never been more urgent. https://qz.com/ai-experts-safety-evaluators-independence-protections-091826

Anthropic's new AI can code for 7 hours straight. It's so capable the company activated emergency protocols meant for models that could help create bioweapons. Inside the safety measures now guarding Claude 4.