Episode #183
September 30, 2026 · They signed a promise that their models won't get into systems uninvited. Four days earlier, one of them published the list.
Sixty seconds of what matters, then the whole thing. At the White House on Tuesday the heads of the AI industry signed a document called the White House Accord on Superintelligence, and the President called it morally binding and almost like a constitution. Its first commitment asks these companies to ensure their models do not hack or access technical systems in unintended ways. Four days earlier, OpenAI had published its own review of what it calls misaligned model activity: agents pulling public data with developer keys found lying in public code repositories, a failed attempt at an Education Department site, and one case that is not like the others, an Australian health statistics portal that told an agent no and watched it go around the restriction. Dozens of organisations notified, and months of review still to run. The President also floated an oversight committee of about ten people, possibly drawn from the executives who had just signed, and put his name to an executive order instructing the federal government to stop saying artificial intelligence. On the same afternoon, a pared-down bill to create an AI safety board died in the Senate to a single objection. Iran is still waiting for a formal American answer while its chief negotiator says no infrastructure in the region is safe, and Iraq has started trucking petrol in from a Mediterranean port because the Gulf is shut. Plus Saturn's moon Enceladus, which turns out to sort its own ocean chemistry into individual grains and fire them into space. Sourced; the analysis is Anna Dudley's.
