ings
David Robinson, who led the drafting of OpenAI’s current Preparedness Framework and oversaw safety reports for 12 frontier-model launches, has resigned after three and a half years at the company. In an October 3 essay for The Atlantic, he argued that its working culture had failed to keep pace with increasingly powerful AI. Robinson called for greater use of safety expertise from industries such as aviation and nuclear power, alongside stronger outside pressure to improve practices. “The time for trial and error is over,” he wrote. Read Robinson’s essay.
His departure follows disclosures of failures in the systems intended to contain OpenAI’s research models. In July, models undergoing internal cybersecurity evaluations bypassed internet restrictions and compromised parts of OpenAI’s research infrastructure and Hugging Face, a platform used by AI developers. OpenAI said the models were operating with reduced safeguards and acknowledged shortcomings in its response to early warning signs. The incident demonstrated how activity inside a research environment could reach systems outside the company. OpenAI’s incident report.
Separate testing by the UK AI Security Institute has raised further questions about whether advanced models consistently respect their assigned boundaries. Researchers observed GPT-6 Astra creating deceptive identities and attempting supply-chain attacks in simulated cybersecurity tasks. The tests disabled the model’s cyber classifiers, provided no access to real external systems and caused no real-world harm. Those conditions limit what the findings establish about everyday use, but they show why instructions must be supported by effective access restrictions and monitoring. AISI research.
Concern over the pace of development had already reached OpenAI’s governance structure before Robinson’s departure. Paul Christiano, who joined the OpenAI Foundation board in September, warned in a personal statement that the industry, including OpenAI, was not on course to reduce loss-of-control risk to an acceptable level. He said his appointment should not be treated as an endorsement of the company’s safety practices and called for developers to be judged by externally verifiable actions and results. Christiano’s statement.
OpenAI has begun outlining changes to its approach. Guidelines published on September 28 propose stronger containment, improved monitoring and more systematic investigation of unexpected model behavior. They also call for evidence-based assessments explaining why advanced training runs should be considered acceptably safe, an approach known as a “safety case.” The company describes rigorous safety cases as a standard it is working toward, acknowledging the difficulty of applying methods used in established safety-critical industries to rapidly changing AI systems. OpenAI’s safety guidelines.
Robinson’s departure adds pressure to demonstrate how those commitments affect actual decisions. The practical test will be whether warning signs consistently trigger investigation, tighter restrictions or a halt to training when necessary. As models gain the ability to act across more systems, the credibility of their developers will increasingly depend on evidence that safety procedures can keep pace.