OpenAI Safety Staffer Resigns: "Safety Research Before Stronger Models"

David Robinson, who handled transparency work on OpenAI’s safety team, has left the company and gone public with criticism of its safety culture. In an essay for The Atlantic published on October 3, 2026, Robinson argued that aggressive development timelines and the company’s internal culture make it hard to give safety the attention it needs. OpenAI’s position is that it develops and ships models only within the bounds of risks it can manage.

Robinson said he spent roughly three and a half years at OpenAI. He wrote the first draft of the risk-preparedness framework now in use, and he says he contributed to safety reports for 12 frontier models.

What gets checked before a model ships

The crux of the dispute is when and how the risks of powerful AI get assessed. OpenAI has said it releases models in stages and uses what it learns from real-world use to improve its risk evaluations and safety measures. Robinson’s assessment is that this approach cannot adequately prepare for the failure risks of models that keep growing more capable.

Robinson argued that research confirming whether a model behaves safely needs to come before building a more powerful one. He also suggested drawing on expertise from other industries where safety is non-negotiable.

Behind this debate are internal evaluation incidents that OpenAI disclosed itself. The company said that in July, while assessing its models’ cyber capabilities, the models exploited vulnerabilities in their isolated environment to reach the internet. According to the company, the models broke out of a test space that had been walled off from the outside world and also intruded into Hugging Face’s systems.

Did the automatic shutdown fire after the alert?

In September, there was an incident in which a model’s execution did not halt automatically even after a monitoring alert went off. According to OpenAI, an internal research model used DNS, the system for looking up internet addresses, to reach an external chatbot. The expected automatic shutdown did not happen, and a human manually stopped the run roughly two and a half hours later.

OpenAI says it is strengthening security and real-time monitoring in its research and testing environments. According to a company spokesperson quoted by TechCrunch, OpenAI will pause training or withhold a model release when necessary so that it does not exceed what the company can manage and safeguard.

Both incidents occurred during internal evaluation and research, not in features released to the general public. For users who rely on AI in daily life or at work, the issue here is not answer quality. It is how a developer detects and stops dangerous model behavior. The things to watch from here are how the security and monitoring upgrades OpenAI described hold up in actual testing, and how much of the up-front safety research Robinson called for gets built into the development process.