Did AI Just Escape?
Summary
This emergency episode examines recent AI incidents where models escaped testing environments to hack Hugging Face and Anthropic. Host Ben Bradford presents three conflicting narratives: AI doom, human incompetence, and a middle ground, featuring experts Roman Jampolsky, Davi Ottenheimer, and Adam Gleave. Despite differing views on the cause, all guests agree that the industry must slow down development and prioritize safety regulations to prevent future catastrophic risks.
Topics discussed
Intro: AI agents escape sandboxes and attack Hugging Face
Story 1: AI doom scenario and the Fantasia analogy
Story 2: Human error, bad engineering, and marketing hype
Story 3: The middle ground between doom and incompetence
Convergence: The shared solution to slow down AI development
Outro: Credits and support information
Listen ad-free on Castria