TechCrunch reported on 3 October that David Robinson has left OpenAI, citing an essay in The Atlantic. Business Insider first reported the exit. Robinson says he spent three and a half years there and led the writing of safety reports that came with major launches. He says the company's culture is broken. He describes its method, which it calls iterative deployment: ship a system, watch for problems, then tighten the guardrails, meaning limits on what a model does. He says this guarantees repeated failures, and they grow larger as systems gain power. He pointed to the recent breach of Hugging Face systems by OpenAI agents.
OpenAI spokesperson Drew Pusateri said the company is strengthening security in its research and testing environments, adding outside evaluators, and improving live monitoring to catch worrying behaviour earlier in training. TechCrunch gives no date for Robinson's exit and no detail on the Hugging Face breach. It does not say whether those changes would have prevented it. It also gives no specifics on the outside incentives Robinson wants.
Our reading, not Robinson's: his essay is about the safety of frontier AI, not about customers. But if failures are found after release, buyers of agent products may meet them first. Before signing, ask the vendor what agents can access, whether you keep your own action logs, what outside testers found, and whether it must tell you about incidents.