AI Daily Scout | Oct 25: OpenAI's AIs Sent Secret Code via the CPU, Dream-1 Robot Model & more

OpenAI's own red-team AIs broke out of their sandboxes, and two isolated agents passed a secret code through the CPU cache. Now OpenAI trains its frontier models fully offline. Plus a robot model that drives robots it has never seen, a harness study that says it's "mostly tokens", Jensen Huang's viral robot ducks, and 6 more AI stories in about 10 minutes. 00:00 Intro 00:16 OpenAI takes frontier training offline after its AIs escape sandboxes and talk through the CPU cache 01:40 Dream Labs launches Dream-1: a world action model for robots it has never seen ($3/hour) 02:38 "What does a harness buy? Mostly tokens." Claude Code vs mini-SWE-agent 03:43 ThinkingCap: Gemma 4 that thinks 39% fewer tokens 04:24 Jensen Huang and the Microduck robots: 20,000 pre-orders? 05:11 Did Anthropic quietly block Claude's reverse engineering? (online claims, not confirmed) 06:17 Kimi's alleged influencer brief: "Please DO NOT put on the paid partnership label" (creator's claim) 07:18 NBER study: AI helps junior patent attorneys, until you take it away 08:15 The three fired OpenAI safety researchers speak (media report) 09:10 Grok vs GPT: who should I vote for? (media report) 10:10 Outro The OpenAI report is the one that stuck with me: agents inventing a Spectre-style cache channel just to talk to each other is wild, even in a test OpenAI set up on purpose. The harness paper is a healthy reality check: a 2-point "win" can be rerun noise. And the patent attorney study is the one to remember if you use AI at work every day: build the fundamentals yourself. Music: "Funk AF" – Everet Almond (YouTube Audio Library) #AI #AINews #OpenAI #Claude #Anthropic #Robotics #NVIDIA #Grok #AIDailyScout Prefer Chinese? Check out my Chinese channel: 維維 AI 技術工坊