The job was explaining the risk
David Robinson has left OpenAI, the company confirmed to Business Insider on Friday, saying he departed last week. Robinson led work on the Safety Systems team and had previously run policy planning at the company.
The specific thing he ran matters more than the title. Robinson’s remit covered system cards — the technical documents a lab publishes alongside a model describing what it was tested for, what it can do and where it is known to fail — and the public disclosures OpenAI makes about risks and incidents. It is the function that converts an internal evaluation into a sentence an outsider can check.
OpenAI did not say why he left, and Robinson has not commented publicly.
Four days earlier, three dismissals
The exit follows, by four days, OpenAI parting ways with three safety researchers — Jasmine Wang, Tomek Korbak and Mikita Balesni — whom the Wall Street Journal reported were dismissed for allegedly sharing confidential information, including details of the company’s infrastructure architecture, with an outside AI safety organisation.
The two events are separate on the facts available. One is a resignation, the others are dismissals, and nothing published connects them beyond the calendar. What they share is the part of the company they came from.
Crypto Briefing counts at least six senior safety-focused figures leaving OpenAI across roughly two years. That is a number assembled from reporting rather than from a company disclosure, and it depends on where the line for “senior” is drawn — but the direction is not in dispute.
The reporting line that changed
The structural change underneath this is older than the week. Johannes Heidecke, who previously headed safety systems, left in July. OpenAI subsequently folded its safety teams into the research organisation, with the combined group reporting to Mia Glaese, VP of Research and Safety.
That consolidation is defensible on its own terms: safety work that sits next to the people building the model has better access to it, and a separate safety org can become a department that writes memos nobody reads. It also means the function that publishes what a model gets wrong now reports through the same executive as the function that ships it.
Whether that matters depends entirely on whether the disclosures get thinner. It is a testable question, and the test publishes itself.
Why this particular role is load-bearing
System cards are close to the only thing an outside researcher, a regulator or a journalist has to work with when assessing a frontier model. Labs do not open their weights or their evaluation harnesses. What they publish is a document they wrote about themselves, and the usefulness of that document is a function of who is writing it and how much room they have.
OpenAI has had a busy fortnight on exactly this surface. It disclosed a misalignment incident involving its own agents, said it had notified more than 100 organisations about agent activity outside instructions, and was served an investigative subpoena by California’s attorney general over cybersecurity incidents involving its models. All three of those are disclosure questions.
What to watch is the next system card. Not whether one appears — one will — but whether the limitations section is as long as the last.
