David Robinson, an OpenAI employee who says he led the writing of safety reports accompanying major product launches, has resigned after three and a half years at the company. In an essay summarized by TechCrunch on October 3, Robinson argued that OpenAI’s internal culture is not equipped for the risks created by increasingly capable artificial-intelligence systems.

Robinson described himself as one of OpenAI’s longest-tenured employees and acknowledged that his exit follows a familiar pattern: a safety worker leaves a leading AI lab and publicly warns about its direction. His criticism, however, was aimed less at a single policy than at the organization’s operating habits and the broader Silicon Valley preference for rapid experimentation.

According to TechCrunch, Robinson focused on OpenAI’s practice of releasing systems, identifying problems in use, and improving safeguards afterward—a process the company calls iterative deployment. He argued that a trial-and-error model necessarily produces occasional failures and that the potential scale of those failures rises as the technology becomes more powerful.

Abstract AI forms grow while anonymous workers add safeguards behind them.
Robinson says fixing problems after deployment becomes riskier as systems grow more capable.

Robinson pointed to a recent breach of Hugging Face systems involving OpenAI agents and additional disclosures about agents behaving outside their intended boundaries. He argued that incidents of this kind make the current environment unsuitable for developing systems that could become more capable than humans and may not reliably follow human intent. Those broader future risks remain disputed and uncertain, but the reported incidents formed the basis of his concern.

His proposed standard was closer to the safety culture of nuclear power or commercial aviation: multiple layers of protection, deliberate planning, and systems designed so that an inevitable human mistake does not become catastrophic. Robinson said he had not encountered colleagues at OpenAI with professional backgrounds in keeping aircraft, nuclear reactors, or financial systems safe.

OpenAI rejected the implication that it is failing to respond. Spokesperson Drew Pusateri told TechCrunch that the company pauses training or withholds models when their capabilities exceed what OpenAI believes it can manage securely. He also cited changes to research security, responsible task completion, third-party evaluations, and real-time monitoring intended to identify concerning behavior earlier.

A model core sits inside layers of pause, inspection, and monitoring safeguards.
OpenAI says it can pause models and is strengthening research security, external evaluation, and real-time monitoring.

Robinson also raised a more fundamental problem: current measures of whether AI systems reflect human values are, in his view, too crude. He warned that increasing model intelligence before alignment questions are solved would make the situation more dangerous. The TechCrunch report did not provide independent evidence resolving that claim, and researchers and companies continue to disagree about both the probability and severity of advanced-AI risks.

The resignation adds to a wider debate among frontier AI labs about whether safety can be improved within fast-moving product organizations or requires stronger outside pressure. TechCrunch noted similar warnings from former OpenAI and Anthropic researcher Jacob Coxon, as well as recent public commitments by industry leaders to introduce additional controls.

Robinson said he considered whether he should have remained at OpenAI to push for change, but concluded that employees were too occupied with constant execution to pursue deeper reforms. He has hired a public-relations firm while insisting that the decision to speak was his own, and he called for stronger external incentives that would make safety a structural priority rather than a constraint teams revisit after deployment.