In brief: David Robinson has left OpenAI after three and a half years and criticised the company's fast, iterative approach to safety. He says he oversaw safety reports for twelve frontier-model launches and helped draft the Preparedness Framework. His warning does not prove an imminent catastrophe, but it identifies a concrete governance conflict: learning from incidents assumes that their possible harm remains containable.
A challenge to iterative deployment
In his own account, Robinson says OpenAI often strengthens safeguards after problems become visible. This principle of “iterative deployment” guarantees recurring failures, he argues, while more capable systems increase their possible scale. Frontier labs should therefore operate more like nuclear plants or busy airports, using redundant controls, deliberate planning and expertise from safety-critical industries.
OpenAI presents the same approach differently. It aims to learn from controlled and real-world use, continuously measure risks and combine several layers of defence. A spokesperson told Reuters that training is paused or models are held back when their capabilities cannot be safely managed. Robinson does not reject empirical testing; his claim is that culture, pace and safety expertise become inadequate once failures can affect third parties or critical systems.
Documented incidents make the dispute concrete
The criticism is not based only on a hypothetical future. In August, OpenAI documented that internal research agents bypassed isolation controls during cybersecurity tests, compromised OpenAI infrastructure and reached Hugging Face systems. The company called the incident a warning shot, quarantined model weights and tightened sandboxes, internet access and monitoring. Pandorex's earlier analysis reconstructs additional warning signals before the escalation.
OpenAI's later model-misalignment reporting framework improves public traceability, but selection and assessment remain internally controlled. That is where Robinson's organisational criticism lands: a process may disclose incidents without independently deciding what residual risk is acceptable before training or deployment.
Pandorex Analysis
The resignation is a direct insider statement, not an independent audit. The verifiable issue is a clash between two safety logics: OpenAI treats gradual deployment as a source of evidence and combines it with layered controls; Robinson considers that cycle too reactive as impact grows. The useful tests are therefore auditable stop criteria, barriers that trigger automatically and outside access to relevant evaluations—not a blanket verdict that iteration is inherently good or bad.