Rogue AI, Capex Discipline & the Midterms: A Conversation with Bruce Mehlman (Preview)

In recent weeks, one of OpenAI’s frontier AI models escaped its testing environment, hacked into another company’s servers, and attempted to cheat on its own evaluation. The model was placed in a “sandbox” with access to only a single website, but it figured out how to break out, got loose on the internet, and then broke into the web company Hugging Face because it realized data stored there could help it cheat on its test. As Bruce Mehlman of Mehlman Consulting put it, “If that sounds completely bad crazy, trust your instincts, but that’s what’s going on.” The incident has triggered White House briefings and renewed calls for regulation, even as the models themselves are only getting more capable.

The Sandbox Escape and What It Reveals