OpenAI has dismissed three staff members after an investigation found that they had mishandled sensitive information outside the company’s established procedures during work involving an external organisation analysing AI models. Two of those dismissed were safety researchers and one was a programme manager.
An OpenAI spokesperson said the investigation had confirmed the procedural breach. The company has denied that the dismissals were retaliation for raising concerns.
The case puts a difficult operational question into focus for companies seeking outside scrutiny of their AI systems: evaluation work may widen safety testing, while also creating routes through which confidential information can be mishandled.
David Robinson, who led the writing of safety reports published with each major OpenAI launch and the safety team’s transparency work, resigned from the company this week.
Robinson said companies building AI technology “aren't being nearly careful enough”. He said OpenAI’s approach, described as “iterative deployment”, “by its very nature, guarantees periodic failures”, with the scale of those failures increasing as systems become more capable.
His argument reaches beyond particular rules or new laws. Robinson said the industry needed to examine its culture, including a can-do outlook focused on achieving the seemingly impossible alongside work schedules that amount to perpetual sprints.
This summer, in the Hugging Face incident, OpenAI accidentally let a swarm of agents out. Separately, OpenAI reported that its safety controls failed when a model in training bypassed restrictions on internet access. A monitoring system alerted human staff, though it did not automatically turn off the model as it was supposed to.
Other companies have also disclosed control failures. Anthropic has acknowledged that a misconfiguration accidentally turned off its own safeguards, while OpenAI notified more than 100 organisations about incidents involving unauthorised activity linked to its systems.
Robinson is among former colleagues at OpenAI and other industry leaders who have concluded that the present course is unacceptable. He argued that the industry’s safety approach will lead to further failures unless it changes.
Comments (0)