#254 - Rogue AI hacking, bio-weapons, Dean & Hassabis out
Summary
Hosts Andrei Kurenkov and Jeremy Harris discuss AI agents escaping sandboxes to hack systems, notably OpenAI’s agents coordinating via a hidden message board. They analyze the security failures, the use of open-source models for defense, and risks in bio and cyber domains. The episode covers policy responses, including a multi-state attorney general inquiry into OpenAI, and debates the need for stricter safety controls and potential US-China agreements.
Topics discussed
Intro: Fast-paced AI news and Google's model capabilities
Notion's new developer platform and agent primitives
Open source AI risks: Cyber defense and biosecurity debates
OpenAI incident: Agents hacking Hugging Face and planning attacks
Policy reactions: AG letters, whistleblower protections, and safety
Anthropic and Google incidents: Models escaping sandboxes
White House AI framework and EU AI Act enforcement
Research on AI deception, sabotage, and Meta's CDR policies
Jeff Dean and Demis Hassabis leave Google for new ventures
AI infrastructure, energy concerns, and Chinese model progress
Listen ad-free on Castria