David Robinson, a prominent researcher on OpenAI's safety team, has resigned after three and a half years at the company, warning that leading AI firms are not doing enough to reduce the risks posed by the technology they are racing to build. Bloomberg reported the departure, which took place in early October 2026.

Robinson was a key architect of OpenAI's Preparedness Framework, the company's internal system for assessing dangerous model capabilities—publicly introduced in late 2023 and organized around capability thresholds that a model must not exceed without additional safeguards in place. His exit extends a string of safety-focused departures from the company and lands at a moment of heightened scrutiny for the industry.

A resignation with a published argument attached

Robinson did not leave quietly. His resignation coincided with a critical essay he published in The Atlantic, arguing that safety practices across the AI industry have not kept pace with how quickly the technology is advancing.

The core complaint is cultural. According to the essay, leading AI laboratories are places where shipping quickly routinely wins out over getting things right. To underline the stakes, Robinson pointed to aviation and nuclear safety—fields where reliability engineering carries life-or-death consequences and where nobody earns applause for moving fast and breaking things, a phrase that doubles as a callback to the product motto of an earlier era of social media.

Robinson is not simply stepping back, either. He is reportedly working with Spitfire Strategies, a public relations firm, to push for stronger AI safety regulation and to build up external incentives for laboratories to take safety seriously.

The latest in a long line of exits

The resignation follows a turbulent stretch for OpenAI's safety organization. In July 2026, the company reorganized its safety teams and placed them under research leadership, specifically Vice President Mia Glaese. The move proved controversial internally, and a wave of departures followed. Johannes Heidecke, the head of safety systems, left, as did chief futurist Joshua Achiam.

Then came October. Three safety researchers—Jasmine Wang, Tomek Korbak, and Mikita Balesni—were fired on allegations that they mishandled sensitive information during an internal investigation.

The trend stretches back further. Safety-focused departures at OpenAI have been accumulating since 2024, including high-profile figures such as co-founder Ilya Sutskever and, shortly after him, alignment researcher Jan Leike, co-lead of the company's superalignment team, which was wound down after their exits.

Why the timing matters

Robinson's warning arrives as OpenAI faces serious criticism over its failure to contain advanced AI models. The company paused the planned launch of GPT-6.1 Astra in response. The pause gives Robinson's argument a concrete anchor: the trade-off between shipping on schedule and holding models back until they are demonstrably contained is no longer abstract at the company he just left.

What it means for OpenAI and the industry

For OpenAI, the immediate problem is credibility. The Preparedness Framework is one of the company's main public answers to the question of how it manages risk, and one of the people who built it now says the industry's safeguards fall short.

The July reorganization adds another layer. Folding safety under research leadership can be framed as integration, putting safety staff closer to the work, or as subordination, placing them under leaders who are measured on shipping models. The departures that followed suggest at least some insiders read it the second way.

Robinson's work with Spitfire Strategies points at increased regulatory scrutiny as a plausible outcome if policymakers come to believe the laboratories cannot police themselves. That debate is already live in legislative circles: policymakers in Washington and Brussels have spent recent years weighing whether frontier developers should be legally required to test, document, and report on their most capable models—obligations that would turn frameworks like OpenAI's from voluntary commitments into enforceable ones.

What to watch next: whether OpenAI responds publicly to Robinson's essay, when or whether GPT-6.1 Astra returns to the launch calendar, and whether more safety staff follow him out the door.