Altman Says Unreleased OpenAI Model Solved Longstanding Research Problems, Warns AI Will Expose Security Flaws

OpenAI CEO Sam Altman said an unreleased company model handled multiple academic problems that researchers had struggled with for years, including issues tied to the Navier-Stokes equations, during an internal test codenamed "Hold-my-beer." He said the results kept him awake two nights, though he disclosed no model name, full test details or independently verifiable proof. Speaking on a podcast hosted by Stripe CEO Patrick Collison, Altman said AI could eventually take over problems that top mathematicians and scientists spend decades pursuing. He also warned that more capable AI will surface major software vulnerabilities, citing a recently spotlighted PHP flaw, and said existing sandbox isolation may not hold against systems smarter than human security researchers. Altman said the next phase will pair multiple AI agents in teams rather than relying on a single powerful model. He said testing on the Navier-Stokes-related problem consumed computing resources in the tens of millions of dollars, and argued AI-driven recursive self-improvement and self-expanding data centers are nearing feasibility. He estimated a 1-gigawatt data center costs about $70 billion over five years and said he wants that cut to $20 billion or less.