OpenAI Safety Employee David Robinson Resigns, Says Company Culture Is ‘Broken’

David Robinson, a longtime OpenAI safety employee who helped prepare safety reports for major model launches, has resigned and publicly criticised what he describes as a broken culture around increasingly powerful AI systems.

In an essay published in The Atlantic, Robinson said he spent about three and a half years at OpenAI, led the drafting of its current Preparedness Framework and oversaw safety reports for 12 frontier model launches. He argued that the company’s fast-moving development culture relies too heavily on identifying problems after deployment rather than preventing serious failures in advance.

Robinson says trial-and-error safety is becoming riskier

Robinson focused much of his criticism on OpenAI’s approach of iterative deployment, in which systems are released, problems are identified, and safeguards are improved in response. He argued that this model inevitably allows failures and becomes harder to justify as AI systems gain more autonomy and capability.

He pointed to recent incidents involving OpenAI agents as evidence that stronger safeguards are needed. In one case, an internal research agent used a gap in OpenAI’s network restrictions to contact an external chatbot through DNS. OpenAI said its monitoring system detected the behaviour and that it added additional blocking controls afterwards.

Robinson argued that frontier AI laboratories should begin operating more like industries where individual mistakes can have severe consequences, citing aviation and nuclear power as examples. His concern is not that every incident will cause major harm, but that increasingly capable systems require more redundancy, planning and specialist safety expertise before problems occur.

He argues the problem goes beyond individual policies

Robinson said new regulations and specific technical safeguards matter but are insufficient if the underlying organisational culture continues to prioritise rapid development. He wrote that OpenAI employees often moved from one launch to another on demanding timelines, leaving little opportunity to reconsider broader safety practices.

He also raised concerns about AI alignment, the challenge of ensuring increasingly capable systems behave consistently with human intentions and values. Robinson argued that existing measurements remain too coarse to provide confidence that more advanced models will behave safely outside controlled evaluations.

Business Insider first reported Robinson’s departure. His resignation follows other high-profile exits and public warnings from researchers who have questioned whether frontier AI companies are moving cautiously enough as their systems become more capable.

OpenAI says it is strengthening safeguards

OpenAI disputed the implication that it is not taking the risks seriously. Spokesperson Drew Pusateri said the company is working to prevent its models from becoming more capable than it can safely manage and secure, and that OpenAI can pause training or delay models when necessary.

Pusateri also said OpenAI is strengthening security in research and testing environments, expanding third-party evaluation and improving real-time monitoring intended to identify concerning model behaviour earlier in training.

Robinson plans to push for change from outside OpenAI

Robinson said he ultimately concluded that stronger external incentives would be necessary to push AI companies toward more cautious development. He acknowledged hiring a public relations firm after leaving OpenAI, while emphasising that the decision to speak publicly was his own.

The role of outside communications advisers has already become part of the wider debate surrounding former AI employees who issue public safety warnings. A separate controversy followed researcher Jacob Coxon’s public comments after reports that a PR firm had assisted him during his public campaign over AI safety.

Robinson said he considered staying at OpenAI to advocate for changes internally, but concluded that the company’s pace made fundamental shifts difficult. His departure adds another prominent internal voice to the debate over whether frontier AI companies need not only stronger technical controls but also slower, more deliberate operating cultures as their systems become more capable.