A New Plan Aims to Pause AI Before It Gets Dangerous

You've heard scary talk about AI becoming too powerful. But how do you actually stop that? One new proposal takes a page from international arms control—specifically, the deal that kept Iran a year away from building a nuclear bomb. The trick isn't banning AI entirely. Instead, set a limit that keeps any AI system far enough from the "danger zone" that if someone tried to cross it, we'd have time to catch them.

Why does that matter to you? Because the risk isn't just a robot uprising. The danger is AI being used by a small group to create bioweapons, cyberattacks, or other mass-casualty events. The author argues that if we freeze AI at, say, 80% of the way to that danger point, the remaining 20% could be finished much faster by rogue actors—with AI's help. So the threshold needs to be like a tripwire set early enough to give us warning.

The proposal borrows the idea of "breakout time": requiring that it take a skilled team at least a set amount of time—maybe months—to go from allowed AI capabilities to truly dangerous ones. It also says the limit shouldn't be fixed permanently; we'd need to adjust it as technology evolves. The author admits the plan is rough around the edges, but published it quickly ahead of a major political summit in late 2026, hoping leaders will start talking about enforceable AI limits.