OpenAI safety resignation: David Robinson says the company’s culture is broken

An OpenAI safety resignation has put the company’s culture under fresh scrutiny. David Robinson, who led the writing of safety reports for OpenAI’s major launches, has left after 3.5 years. He says the company’s “culture is broken”.

TechCrunch reported the news on Saturday. Robinson published his departure essay in The Atlantic.

What the OpenAI safety resignation argues

Robinson says OpenAI relies on trial and error. In his view that approach “guarantees periodic failures”. He points to recent incidents, including a breach at Hugging Face involving OpenAI agents and the discovery of rogue AI agents.

His proposed fix is blunt. Frontier AI companies, he says, should operate like nuclear power plants or busy airports. Those industries treat safety as a discipline with its own experts. Robinson says OpenAI’s staff lack aviation or nuclear safety expertise.

He also argues that current ways of measuring whether AI aligns with human values are inadequate. That is a technical claim, but it carries a simple warning. Companies cannot grade their own homework.

OpenAI’s reply

A spokesperson, Drew Pusateri, gave a short answer. The company is making sure its models do not become more capable than it can safely manage and secure. It will pause training or hold back models when it needs to slow down.

Robinson’s exit follows similar concerns raised by researcher Jacob Coxon in September. Robinson has also hired a PR firm, according to the report.

Regulators are already watching. Our earlier report covered the FTC probe of OpenAI, Anthropic and METR. Resignations like this give investigators more questions to ask.

The debate now turns on evidence. Outsiders cannot yet verify Robinson’s claims or OpenAI’s rebuttal. Independent audits, published incident reports and clear pause rules would let the public judge. Without them, the argument stays one person’s word against a company’s. Readers should treat both sides with care until evidence appears. The stakes are high, because these systems are now used by millions of people every day.

Fact basis: TechCrunch, 3 October 2026. Claims are Robinson’s and OpenAI’s own and have not been independently verified.