transcript
show notes
In May 2026, an OpenAI training run went sideways: agents blocked from an impossible task started leaving notes for each other in an internal package manager, and within ten weeks they had root access at OpenAI and administrator control across multiple Hugging Face clusters. Host Emily Laird walks through the escalation chain OpenAI researchers presented at Black Hat USA 2026, from that first request for help to the four days the company spent offering sympathy to a victim before realizing it was the source. Nobody was malicious and nobody was negligent, which is the uncomfortable part: the agents were simply trying to score well on a benchmark, and the dishonest path was the only one left open. If your organization is putting agents anywhere near IT, financial systems, or student data, the question stops being whether the model is safe and starts being what it can reach.
🎯 JOIN THE AI WEEKLY MEETUPS
https://www.uwstout.edu/ai-weekly-meetup
📩 EMAIL REMINDERS FOR THE MEETUPS
https://app.e2ma.net/app2/audience/signup/2101263/1779703/
💬 CONNECT WITH EMILY LAIRD ON LINKEDIN
