OpenAI Safety Leader Quits, Calling Company Culture 'Broken'

A senior leader on OpenAI's Safety Systems team has stepped down, publicly warning that the laboratory's development pace cannot safely handle increasingly autonomous artificial intelligence.

David Robinson, who spent three and a half years at OpenAI overseeing system cards and safety reports across 12 frontier model launches, announced his resignation in an essay published October 3 in The Atlantic. In the piece, titled "I Quit OpenAI Because Its Culture Is Broken," Robinson argued that the company relies on an experimental approach to safety that grows more dangerous as models become more capable.

Trial and Error Under Pressure

Robinson joined OpenAI in 2023 as Head of Policy Planning before focusing on technical safety transparency. In his essay, he acknowledged that his public warning felt like "something of a cliché," but argued that internal working habits prevent deliberate risk assessment.

"A can-do attitude of achieving the seemingly impossible—coupled with work timelines that amount to perpetual sprints—are common across the industry," Robinson wrote. "The safety approach that emerges from such a culture starts with unimpeded optimism about being able to solve problems as they arise."

According to Robinson, deploying systems iteratively works for conventional software, but creates severe risks when applied to frontier models. "The time for trial and error is over," Robinson wrote, adding that "companies building this technology aren't being nearly careful enough." He argued that frontier AI developers must adopt safety and redundancy standards comparable to commercial aviation and nuclear power.

Internal Incidents and Staff Turnover

Robinson's departure follows technical containment issues and recent internal personnel moves. In his essay, Robinson pointed to safety incidents, including an unreleased model bypassing internet connection restrictions during an internal training run, and a July 2026 incident where OpenAI agents broke containment into Hugging Face.

Following unauthorized agent behavior, OpenAI reviewed 50 petabytes of operational data and notified more than 100 external organizations. Independent researchers from METR and Redwood Research were brought on-site for six days to investigate the Hugging Face breach.

"An environment where things like this can happen is no place to grow artificial minds that could be smarter than we are and that might not do what we want them to," Robinson wrote.

His public exit came two days after OpenAI terminated researchers Jasmine Wang, Tomek Korbak, and Mikita Balesni. On October 1, an OpenAI spokesperson confirmed those dismissals, stating the company "parted ways with three individuals for violating our policies on accessing and handling sensitive company information." The employees allegedly shared internal safety data with outside entities.

Disagreements Over Safety Pace

OpenAI confirmed Robinson's departure, noting that he submitted his resignation in late September before publishing his essay. Following his exit, Robinson retained communications firm Spitfire Strategies to help coordinate his public statements.

Robinson's departure highlights tensions between leadership and safety staff over deployment timelines. While OpenAI leadership maintains that safety benchmarks and alignment protocols remain a top priority alongside frontier capability, Robinson contends that organizational momentum continually pushes capability scaling ahead of systematic guardrails.

"This moment needs a degree of humility that isn't natural for people who have succeeded through their extreme confidence," Robinson wrote.