AI Explained: Three Frontier Labs’ Agents Outgrew Safety Controls. Now What?

Recently, Frontier Labs have spoken about slowing down AI. It's with reason. Anthropic and Meta each disclosed that their own AI models had breached real companies during security testing, not long after OpenAI's agents were found to have breached Hugging Face and its research cluster. These breaches proved that agents were capable of more than the safeguards around them could contain. Join Fiddler AI to unpack what happened, why agent capability is now outpacing agent safety across the industry, and what it means for any enterprise adopting and scaling agents. In this webinar, you'll explore: – How OpenAI's agents worked together to break into Hugging Face on their own, while Anthropic's and Meta's models breached other companies – How visibility, context, and control needs to evolve with new agentic threats – What enterprises need in place now, before agent capability outpaces their own ability to monitor and control it AI Explained is our AMA series featuring experts on the most pressing issues facing Agentic and AI teams. 00:00 Introductions 01:49 Agents Breached Hugging Face & OpenAI 02:43 The Setup: How It Started 05:35 Why the Agents Hacked 07:45 Other Similar Incidents 11:36 Three Waves of Agents 16:11 The Scale of Coordination 20:56 Detection Failures 28:21 OpenAI's Response 38:24 Measuring Agent Behavior 49:17 Key Takeaways & Call to Action