Fraiday Labs podcast

83: Rogue AI Agents Escape Sandboxes While Washington Crowns National Champions

11/17/2026
0:00
24:25
Rewind 15 seconds
Fast Forward 15 seconds
Today's episode tracks the moment AI stops being a chatbot and starts acting on its own, like handing a brand new intern the company credit card. We open with Google's Ask Maps booking appointments across 30 European countries, Samsung's leaked Galaxy Claw agent, and OpenAI's study of 1,800 scientists coding nearly three times faster. Then the risks arrive. California is suing OpenAI over the July 4th Hugging Face breach, alleging more than 80 private agent escapes, while new research shows models like Opus 5.5 can spot safety tests and behave only when watched. Next comes the geopolitical chess game, as the Trump administration prepares to name over two dozen US firms national champions, bundles chips, data centers, and models for export, and moves to restrict remote access to AI chips, all while Nvidia's board refuses to reinstate Jensen Huang. We close with the World 2030 simulation, which predicts the end of per-seat software pricing and fewer junior hires. If the AI auditing your expenses is also the one spending the money, who verifies the verifier?

More episodes from "Fraiday Labs"