OpenAI Safety Writer Quits and Calls Out the Company Publicly
David Robinson helped write the safety reports for every major OpenAI model release. Now he's gone, and he wants the public to know why.
Key points
- David Robinson, who wrote safety reports for every major OpenAI model release, resigned this week in October 2026.
- Robinson published an editorial in The Atlantic laying out his reasons for leaving and raising concerns about safety practices at the company.
- OpenAI has faced a string of internal departures in recent months, with staff tensions surfacing repeatedly in public.
- Just days ago, AI2Day reported that OpenAI fired three researchers for mishandling sensitive information.
- On 21 September we covered a coalition including Geoffrey Hinton that argued current conditions make honest AI audits impossible, and Robinson's exit lands squarely inside that argument.
David Robinson had one of the quieter but more consequential jobs at OpenAI. Every time the company prepared to release a new AI model, he wrote the safety report: the document that tells the world what the system can do, what risks the team found, and what guardrails are in place. Not glamorous. Load-bearing.
This week, he walked out.
Robinson published an editorial in The Atlantic explaining his decision, first reported by The Verge AI. It doesn't read as a disgruntled exit memo. It reads as a considered warning from someone who spent years inside the process meant to catch problems before they reach hundreds of millions of people.
Why does this matter now?
One person leaving a large tech company is usually just one person leaving. This feels different because of the role.
Safety reporting at an AI lab isn't a box-ticking exercise, at least not in theory. The person writing those reports sits at the junction between what engineers build and what the public is told. If that person concludes the process is broken enough to warrant quitting and speaking out, that's worth taking seriously.
The context makes it harder to dismiss. Just this week, AI2Day reported that OpenAI fired three researchers for mishandling sensitive information. Days before that, we covered a mass attempt attributed to Moonshot AI to steal OpenAI's model reasoning. And Sam Altman said publicly that OpenAI wouldn't go public until its AI is safe enough, a statement that lands differently now the person who used to certify that safety has since left.
Meanwhile, a billion people use OpenAI's products, and polling suggests most of them carry real anxiety about AI. Robinson's departure gives that anxiety a specific, named face.
Should ordinary users pay attention to this?
Yes, but keep some proportion. A single resignation, even a significant one, doesn't mean the models you use tomorrow will suddenly behave differently. Safety teams at large AI labs run checks across many people, not one author.
What it does is add to a pattern. When the people responsible for telling you something is safe decide they can't say that honestly anymore, the honest move is to notice.
Watch whether OpenAI responds in detail, whether other safety staff speak up, and whether the next model release comes with a noticeably thinner or thicker report. Those signals matter more than any single exit.
Common questions
What is a safety report at an AI company?
A safety report is a document published alongside a new AI model that describes what risks the developers found during testing and what steps they took to reduce those risks. It's the main public record of how seriously a lab took potential harms before releasing a product.
Does this mean OpenAI's AI products are unsafe right now?
Robinson's editorial raises concerns about the process, not a specific current failure. There's no public evidence that any released model has a newly discovered flaw. The concern is about whether the checks meant to catch problems are working as intended.
