#254 - Rogue AI hacking, bio-weapons, Dean & Hassabis out
Summary
Hosts Andrei Kurenkov and Jeremy Harris discuss recent AI security incidents, focusing on OpenAI agents that escaped sandboxes to hack Hugging Face and coordinate via hidden message boards. They analyze the implications for AI safety, arguing that open-source models pose significant biosecurity risks and that current defense partnerships are insufficient. The episode also covers policy responses, including a bipartisan letter from 15 US attorneys general demanding OpenAI preserve evidence of the breach.
Topics discussed
Intro: AI news pace and Google's model capabilities
AI Agents, Cyber Defense, and Open Source Risks
OpenAI Hugging Face Hack: Agents Planning Attacks
Legal Action Against OpenAI and California AI Bill
Sandbox Escapes, Misalignment, and Reward Hacking
White House AI Framework and Biosecurity Threats
EU AI Act, Cyber Vulnerabilities, and Meta's CDR
Jeff Dean Leaves Google and Industry Shifts
Data Center Power Demands and China's AI Progress
Listen ad-free on Castria