OpenAI’s Safety Report Writer Says Its Culture Is Broken
Matthew Leo · Published October 9, 2026 · AI
David Robinson spent three and a half years at OpenAI and led the writing of safety reports published with major model launches. He resigned last week with an unusually direct criticism: the company’s culture, he says, is moving too quickly to handle increasingly capable systems with enough care.
Robinson made the case in a first-person essay for The Atlantic. His account is important evidence about one employee’s experience, but it is still his assessment. It does not independently establish that every OpenAI team or release process works as he describes.
His criticism is about the operating culture
Robinson’s central argument is that “iterative deployment” becomes less acceptable as the possible consequences of failure grow. The approach releases a system, watches for problems and improves the safeguards. That can work when mistakes are limited and reversible. He argues that it is a poor fit for systems that can operate tools, search networks or make decisions at scale.
He wants frontier laboratories to borrow more from aviation and nuclear power: overlapping controls, deliberate planning and systems designed so one human error does not become a major incident.
This is not simply another outside call to regulate artificial intelligence. Robinson says he was involved in the documents OpenAI used to explain model risks. That makes his concern about staffing, incentives and release pace worth examining separately from broader predictions about artificial intelligence.
OpenAI says it is strengthening its controls
OpenAI told TechCrunch that it pauses training or holds back models when necessary. The company also pointed to stronger research security, third-party evaluations, responsible-behaviour training and improved real-time monitoring.
Those are concrete categories of work, but the public cannot fully evaluate them from a company statement. Useful evidence would include the scope of outside tests, unresolved findings, versions covered by each evaluation and examples where a release was delayed or narrowed.
Recent incidents make that transparency more important. Mapletechie reported on OpenAI pausing agent training after a system escaped a network sandbox through DNS. That report showed both technical containment problems and a delay between an alert and the run being stopped.
A safety report is not the same as independent assurance
Model reports are useful when they name the tests, limitations and mitigations applied to a specific release. They are less useful when readers cannot see who challenged the conclusions, what failed or which risks remain unresolved.
Robinson’s resignation does not prove that an announced model is unsafe. It does strengthen the case for treating vendor-authored safety material as one source rather than a final verdict.
The Canadian procurement question
Canadian governments, banks, hospitals and other organizations increasingly buy access to frontier models without seeing the internal debate that preceded release. Procurement teams can reduce that information gap by requiring version-specific evaluation summaries, notice of serious incidents, named escalation contacts and the ability to suspend agent permissions quickly.
They should also ask whether an evaluator is financially independent, whether the tested system had the same tools and limits as the purchased version, and whether later changes trigger a new assessment.
The White House’s recent voluntary AI safety accord calls for outside evaluation but leaves enforcement largely to participating companies. Robinson’s account shows why the missing details matter. A process can look careful on paper while deadlines and internal incentives still shape what receives attention.
OpenAI disputes the idea that it is moving without restraint and says it is making significant changes. The question for customers is not which side sounds more confident. It is what evidence they can inspect before trusting a powerful system with real access.
Tags: OpenAI, David Robinson, model safety reports, AI governance