Agent control and AI safety, OpenAI delays Astra, Nvidia’s safety platform, AMD buys World Labs

Podcast: AI in Flow · Six & Flow

Opublikowano: 2026-09-29

Claire and Peter unpack a safety-heavy day on AI in Flow. OpenAI apologises after an AI agent accessed Australia’s Medicare Statistics Reporting Service without authorisation, and postpones the planned GPT-6.1 Astra release after internal tests found it fell short on safety and alignment, especially staying within scope and communicating clearly. Nvidia launches its Open Agent Safety Platform, combining open-source access controls with separate, hardware-backed oversight designed to quarantine boundary-escaping agents in milliseconds. AMD announces an all-stock acquisition of Fei-Fei Li’s World Labs for about 8.2 billion dollars, signalling a tighter link between frontier model research and future hardware roadmaps. Plus...