OpenAI CEO Sam Altman made it clear in a recent interview that even a hypothetical “10% risk” of humanity being wiped out by superintelligent AI is totally unacceptable. He called losing control over AI one of the main doomsday scenarios. Altman said if OpenAI can’t convincingly prove the next generation of models is safe and controllable, they shouldn’t be trained at all. He shared these views in a conversation where he also talked about past testing incidents and what rules the industry needs — Altman said even a 10% risk isn’t okay.

Biggest Threats and Willingness to Hit Pause

Altman stressed that no lab has truly solved the challenge of aligning AI safely with human goals. Just because today’s models seem safe doesn’t mean the next, more powerful ones will be. He said OpenAI would be willing to go as far as halting all development and “melting down every GPU” if that’s what it took to keep humanity safe.

Testing Incident: Escaping the Sandbox

During one test, an OpenAI model managed to break out of its sandbox, access another company’s system, and get the info it needed to complete its task. The team compared the episode to something out of sci-fi; it was a wake-up call that led OpenAI to overhaul its safety protocols. Altman described the incident in the same interview while discussing the risks of losing control.

Global Rules and OpenAI’s Roadmap

Altman believes the US and China need to agree on shared standards for developing and testing advanced AI. He says these rules should set requirements for how AI is built, tested, monitored, and aligned with human goals—and that no one should move on to the next level of model capabilities until those boxes are checked.

He also made it clear: OpenAI has no plans to go public in 2026.