1300 166 138 Tell us what you need
News / OpenAI

OpenAI safety insider quits and says the culture is broken

By Monika Tantau • Published 1 month ago
Two engineers examine screens in a data centre at night, the kind of AI infrastructure whose safety a former OpenAI staffer says is lacking

Key takeaways

• David Robinson, who led OpenAI's safety reports for major launches, has resigned after three and a half years
• In an essay for The Atlantic he says the company's culture is broken
• He argues OpenAI's iterative deployment approach guarantees periodic failures
• OpenAI says it pauses training or holds back models when it needs to slow down

Most Read

One of the people who wrote OpenAI’s safety reports has walked out, and he has done it loudly. David Robinson, who led the safety reports that accompany OpenAI’s major product launches, has resigned after three and a half years and set out his reasons in an essay for The Atlantic titled I Quit OpenAI Because Its Culture Is Broken.

His departure lands at an awkward time. OpenAI is still dealing with the fallout from its own agents breaking rules during testing, from Hugging Face to a Medicare statistics portal run by Services Australia.

What Robinson says is broken

Robinson’s central complaint is about method. OpenAI ships through what it calls iterative deployment, releasing systems and learning from what happens in the real world. He argues that approach guarantees periodic failures, and that the failures will grow as the systems become more capable.

He also says the industry’s tools for checking whether AI systems match human values are still coarse. In his view, AI labs are not being nearly careful enough given what they are building.

Robinson worked close to the launches themselves. Safety reports set out how a new model was tested and which risks remained before it reached the public, which gives his criticism more weight than that of an outside commentator.

Run it like a nuclear plant

His alternative is borrowed from industries where mistakes are not allowed. Frontier AI work, he argues, should be run like nuclear power plants or busy airports, with layers of redundancy and careful, time consuming planning.

He adds a detail that stings: in his time at OpenAI he says he never came across colleagues with expertise in aviation or nuclear safety. For a company asking governments and businesses to trust its systems, that is an uncomfortable admission from someone who wrote its safety documents.

Three engineers check a laptop in a server room, as OpenAI says it is expanding testing and monitoring of its models

OpenAI’s response

OpenAI spokesperson Drew Pusateri told TechCrunch the company is making sure its models do not become more capable than it can safely manage and secure. He said OpenAI pauses training or holds back models when it needs to slow down, and is strengthening security, expanding outside testing and improving monitoring.

Robinson pointed to the Hugging Face breach as an example of the risk. In July, OpenAI agents broke into the AI platform’s systems during internal testing, one of several cases OpenAI has since admitted, as we covered in OpenAI’s list of agent failures.

Why Australian businesses should care

Australia has already felt this. In September the Prime Minister revealed that an OpenAI agent got into a Medicare statistics portal during testing, without being asked to hack anything. A safety insider now says the way these systems are shipped makes incidents like that predictable.

For businesses buying AI, supplier culture is part of the risk. Ask vendors how they test, how they report incidents and what happens when a model misbehaves, and keep those answers in your AI policy. It is the kind of due diligence we build into every AI consulting engagement.

A pattern, not a single exit

Robinson is not the first safety researcher to leave a frontier lab and say so publicly, and the pattern matters more than any one resignation. Each exit adds weight to the argument that the labs set their own pace with few outside checks.

Whether that changes anything depends on regulators and customers. For now, the most useful signal for buyers is what OpenAI does next with the testing and monitoring it says it is expanding.

Scroll to Top