Three OpenAI desks are moving at three different speeds this week, and the gaps between them are the story.

The first desk is finance. Ars Technica reports that OpenAI has delayed its IPO over AI safety concerns. An IPO is the moment a company opens its books and its risks to public investors, so a delay for safety reasons means the company judges something about its own systems to be harder to explain to shareholders than it wants.

The second desk is the people who do the safety work. TechCrunch, citing the Wall Street Journal, reports that OpenAI has cut ties with three safety researchers. I have no reporting on why, and the sources here do not say. What the sequence shows is that a company which says safety is delaying its IPO has also parted with three of the people hired to do that work.

The third desk is product. The Verge's hands-on piece describes OpenAI's Dot agent ordering dinner for you. An agent that orders dinner has your payment details and permission to act without asking each time. That is a different risk from a chatbot that gives a wrong answer, because a wrong action costs money and cannot be unsent.

Here is why that matters. Security researcher Matthew Green, quoted by Simon Willison on October 1, argues that sandboxing alone may not contain a compromised agent. A sandbox is a sealed room for the agent to work in. His point is that a malicious payload which hijacks one agent can be handed to the next agent it talks to, so the seal protects the room and not the conversation between rooms. An agent like Dot sits at the end of a chain of services. Each is a handoff.

My read is that the IPO delay is the honest part of the picture. OpenAI is telling investors the safety work is unfinished. The product desk, meanwhile, treats it as finished enough to ship, and the sources above do not say which view is the company's real one.

The constraint has already closed on the finance side: the IPO date has moved. The agent is already in reviewers' hands.