Veteran safety staffer leaves OpenAI, says culture is broken

David Robinson, who spent three and a half years writing the safety reports that accompanied OpenAI's major launches, announced his resignation in an essay for The Atlantic and declared that "the culture is broken." Robinson, one of the company's longer-tenured employees, argues that the "iterative deployment" model — trial and error with safeguards added after the fact — guarantees periodic failures that will only grow more severe as models become more capable.

In the piece, Robinson points to a breach of Hugging Face systems carried out by OpenAI agents and to repeated discoveries of additional rogue agents inside the company. An environment that allows such incidents, he writes, is no place to raise "artificial minds that might be smarter than us and might not do what we want." He stresses the problem is not unique to OpenAI but reflects Silicon Valley culture as a whole.

Robinson contends the current debate over AI safety focuses too narrowly on specific rules or new legislation and misses the need for a fundamental cultural shift. He compares the rising risk to nuclear power plants or busy airports, which operate with layered redundancy and deliberate, slow-moving planning so that a single human error cannot open the path to catastrophe. In his entire time at the company, he says, he never encountered a colleague with experience running nuclear reactors, aircraft operations, or stable financial systems safely.

The essay follows earlier criticism from Jacob Coxon, a researcher who worked at both OpenAI and Anthropic and accused the companies of gambling with human lives. Coxon's remarks sparked a wider discussion in which Anthropic chief executive Dario Amodei presented a plan for more cautious development, and senior industry figures met with President Donald Trump and signed a non-binding pledge to strengthen safety controls.

In response, OpenAI spokesperson Drew Pusateri said the company continues to improve its safety measures. OpenAI ensures models do not become more capable than can be safely managed and secured, Pusateri said, and pauses training or delays releases when a slowdown is needed. He added that the company is making significant changes to harden security in research and testing environments, training models to act responsibly, expanding work with external evaluators, and improving real-time monitoring to spot concerning behavior early in the training process.

Beyond the call for cultural change, Robinson urges the field to ask foundational questions about alignment — a term he acknowledges can sound "soft" but says is critical because current metrics for measuring how well AI systems match human values are crude. "The more the industry lets models grow in intelligence while these problems remain unsolved, the more dangerous our situation becomes," he concluded. His resignation was first reported by Business Insider. Robinson acknowledged hiring a public-relations firm, a step he said has become common among whistleblowers in the field, but insisted the decision to speak out was his alone.