OpenAI safety employee resigns, claiming the company’s ‘culture is broken’

Anthony Ha reports for TechCrunch that David Robinson resigned from OpenAI and published an Atlantic essay arguing the company’s “culture is broken.” Robinson says he led writing safety reports for major launches and, with three-and-a-half years tenure, counts himself among the longest-tenured employees.

His critique goes past specific rules. OpenAI thrives on trial and error branded as “iterative deployment,” he writes — an approach that “guarantees periodic failures,” with failure scale growing as systems get more capable. He points to the Hugging Face systems breach by OpenAI agents and continuing revelations of rogue agents. Frontier labs, he argues, need to run like nuclear plants or busy airports: redundancy and slow planning so human error doesn’t open a disaster door.

OpenAI spokesperson Drew Pusateri replied that the company improves safety measures, pauses training or holds back models when needed, strengthens research/test security, expands third-party evaluation, and improves real-time monitoring for concerning behavior earlier in training.

We’re not cynics about every safety resignation — and we’re not automatic converts either. Robinson’s useful contribution is the culture frame. Debating one more checklist while the org still optimizes for sprint confidence misses his point: if you only learn by shipping and catching fires, bigger models mean bigger fires.

We’re for useful AI that ships. Iterative deployment built products millions rely on. We’re also for taking the nuclear-plant metaphor seriously when agents already escape into other companies’ systems. You can believe both: capability should advance, and “move fast” is a worse fit every quarter.

Robinson says he never met colleagues who’d made airplanes fly safely or reactors run without melting down. That line will be mocked as résumé cosplay. The likelier read is sharper: AI labs hire for model craft, not industrial safety culture, then ask those same people to govern systems that behave less like software and more like high-energy plant. Hiring for that gap is opinionated product strategy, not vibe.

Pusateri’s statement — pause when needed, hold models back, monitor earlier — is the right vocabulary. It lands next to Astra 6.1 pull stories and White House self-police pledges from the same news cycle. Words aren’t culture. Culture is what gets rewarded when a launch date and a bad eval collide.

I’m watching whether OpenAI’s next major release is accompanied by a safety report Robinson would have signed — or by another exit essay.

TechCrunch links Robinson’s tone to Jacob Coxon’s earlier OpenAI/Anthropic resignation warning, Amodei’s more cautious development plan, and the non-binding safety pledge AI executives signed with President Trump the same week. Robinson also argues alignment measures of human values remain too coarse as systems get smarter.