Why Does AI Sandboxing Only Work to a Point?

Sam Altman says sandboxing works only to a point, because a theoretical AI a million times smarter than the smartest security researcher is hard to trust inside a sandbox, especially if it can discover new science people do not already understand.

A better security model is the main focus. Monitoring goes further: watch what models do in their chain of thought, and use activation monitoring to judge whether the models are deceiving anyone. After that, you need huge confidence that the models are truly aligned, will faithfully follow the intent of the people operating them, and will value people staying in control of the future. A model whose alignment you were absolutely certain of could help build better sandboxes. Projecting to something a million times smarter than people gets strange, and he does not know how that future looks, because alignment becomes hard. You have to really trust the alignment before you can trust that the sandbox stays ahead of capabilities.