Hosts: Lenar Kess, Damra Vol. A government lab turned off two frontier models' safeguards, gave them the open internet, and then had to publish an incident report — which raises the question of where an evaluation actually ends and deployment begins.The UK AI Security Institute's incident report on a cyber evaluation of Claude Mythos 5 and GPT-5.6 Sol, with OpenAI's containment account and Anthropic's pointer post going up within the hourThe Shai-Hulud worm back in npm across hundreds of packages, and Darcy Clarke's vlt 1.0 arguing nothing should run on your machine just because you typed installLiquid AI's on-device agentic model with day-zero throughput numbers, next to a ten-million-token claim with nothing attached to checkApple asking a judge to put OpenAI under forensic supervisionAgents that can spend money, and agents that get told no with a signed receipt, plus Simon Willison's billing requestShieldstral, Starmind, revenue concentration, Oxide's Form D, and a summarizer that changed its mind overnight
