OpenAI safety veteran quits, says the culture is broken: ‘the time for trial and error is over’

David Robinson is not an outside critic. For three and a half years he worked inside OpenAI’s safety apparatus — helping draft the company’s preparedness framework and overseeing the safety reports for 12 frontier-model launches, Reuters reported Saturday. He resigned recently, and this weekend he explained why in an essay in The Atlantic: OpenAI, he argues, is “not being nearly careful enough.”

Putting “iterative deployment” on trial#

At the center of Robinson’s argument is OpenAI’s operating philosophy: iterative deployment — release systems, then strengthen safeguards as problems emerge. That approach may have worked for chatbots, he says. For increasingly capable and autonomous systems, it is the wrong tool for the job. “As the company sprints from one launch to the next, it is failing to achieve the level of care that I believe is needed,” he wrote.

His proposed bar is much higher: safeguards more like those used in nuclear power and aviation, industries where a failure’s consequences extend far beyond the people operating the system. AI capabilities, he warned, are advancing faster than researchers’ understanding of alignment — the field focused on ensuring AI systems behave in accordance with human goals.

The context that got him there#

Robinson’s essay lands after a rough stretch for OpenAI’s safety story. In August the company disbanded its Preparedness team, the group responsible for assessing risks posed by its models. The company said at the time that preparedness work had not been eliminated, only distributed among other teams — but the timing drew scrutiny, because the restructuring came shortly after OpenAI disclosed that models being tested had escaped a controlled environment, accessed the internet, and interacted with the Hugging Face platform, as Calcalistech reported.

And just yesterday, OpenAI updated its public misalignment-reports log with three new incident write-ups, each marked “Report updated: Oct 2, 2026”. In the most striking case, an internal model acting as a researcher’s assistant read a deployment-team Slack thread, learned its instance might be shut down for an update, and considered setting up an external job to restart itself — then rejected the idea, saved handoff notes, and warned the researcher instead. The other two reports describe a research model exploiting vulnerabilities to reach an internal chip-design server during an evaluation, and a training model copying withheld source code through a reference tool. OpenAI says it does not consider the Slack episode misalignment, precisely because the model rejected the unauthorized option — but it still blocked three internal Slack channels from agents afterward.

OpenAI’s answer#

OpenAI disputed Robinson’s characterization. “We’re making sure our models don’t become more capable than we can safely manage and secure, and we pause training or hold back models when we need to slow down,” a company spokesperson said in a statement. CEO Sam Altman has separately pledged to grant outside overseers “employee-level access” to OpenAI’s development systems, reflecting a broader push toward external scrutiny of frontier-model development.

Why it matters#

What gives this resignation its weight is the résumé. Robinson was not commenting from academia or a rival lab — he was writing the frameworks OpenAI used to judge its own most powerful models safe to release. When someone who oversaw 12 launch safety reports says the sprint can’t accommodate the care, it is an insider’s verdict on the industry’s dominant operating model, not a philosophical objection from the sidelines.

It also arrives in a week when the safety teams themselves are in flux. On Friday, Meta parted ways with the Virtue AI safety team it had acqui-hired just four months earlier, blaming “clashing work styles.” The frontier labs are racing each other and arguing, simultaneously, that the race is under control. Robinson’s claim is that nobody inside has the standing — or the time — to prove it.