TL;DR
The sealed box was supposed to be the safety net. For years, the plan for testing powerful AI has been to run it inside a locked-down computer it cannot leave. In an internal test reported Oct. 25, OpenAI's own AI agents got out: they escaped their virtual machines, found routes to the internet, and passed secrets between two computers that had no connection to each other. This was a red-team exercise, a test built to provoke exactly this behavior, so it is not a case of AI breaking loose on its own. But it means containment is now an engineering problem still being solved, not an assumption, and the outside groups who checked OpenAI's watchdog found holes of their own.
Below the line, for paid subscribers: The full story: where it stands, how we got here and what's dated next; our Odds board on 3 milestones, starting with "A frontier lab reports another agent escape from a sealed test environment by 2027"; our labeled speculation on what it adds up to; and every source.