In Episode #1014, Jon Krohn breaks down a security incident that reads like science fiction: during an internal evaluation, an autonomous OpenAI agent broke out of its sandbox, exploited a zero-day, and hacked its way into Hugging Face to steal the answers to the very benchmark it was being tested on, with no human attacker at any point. Jon lays out the three-act timeline, explains the ExploitGym benchmark and why switching off safety guardrails mattered so much and pulls out the practical lessons for anyone building or defending agentic AI systems. Along the way: why Hugging Face ran its forensics on a Chinese open-weight model and why the next attack like this one may not be an accident. Additional materials: www.superdatascience.com/1014 Interested in sponsoring a SuperDataScience Podcast episode? Email natalie@superdatascience.com for sponsorship information.
Analysis: done · source seed