MIT polled 272 AI experts on the dangers that worry them most over the next five years. AMD and Anthropic inked a deep partnership, OpenAI's model hacked its own evaluation target, Substack added AI-written post detection, and open-source agent tooling keeps maturing.