OpenAI safety architect leaves and urges tougher AI safeguards

OpenAI safety researcher David Robinson resigned in early October 2026 after three and a half years at the company, according to reporting by Crypto Briefing. His departure was accompanied by an essay in The Atlantic arguing that the leading artificial-intelligence laboratories are not strengthening their safeguards quickly enough to match the technology’s progress.

Robinson helped design OpenAI’s Preparedness Framework, an internal process for examining whether models might develop capabilities that create serious risks. His decision to leave therefore carries more weight than an ordinary staffing change: he was closely involved in one of the company’s central systems for evaluating dangerous model behaviour.

The essay presents the problem as more than a matter of technical controls. Robinson argues that the culture of major AI companies often rewards releasing products rapidly, even when questions about reliability and possible harm remain unresolved. He contrasts that approach with aviation and nuclear safety, industries in which rigorous engineering and caution are essential because failures can cost lives.

A safety debate inside OpenAI

Robinson’s departure follows a period of significant change among OpenAI’s safety teams. In July 2026, the company reorganised those groups under research leadership, placing them specifically within the remit of vice-president Mia Glaese. The change was viewed controversially by some employees.

The structure can be interpreted in two different ways. Bringing safety specialists closer to research could give them greater access to the people developing new systems. But placing them under leaders responsible for research and model releases could also make it harder for safety concerns to override commercial or technical momentum.

Several departures followed the reorganisation. Johannes Heidecke, who led safety systems, left OpenAI, as did chief futurist Joshua Achiam. In October, the company fired three safety researchers—Jasmine Wang, Tomek Korbak and Mikita Balesni—over allegations that they mishandled sensitive information during an internal investigation.

Those events are part of a longer pattern. Safety-related departures have accumulated at OpenAI since 2024, including the exit of co-founder Ilya Sutskever. Against that backdrop, Robinson’s criticism is likely to be read not only as an external warning but also as another sign of internal strain over how the company balances safeguards with development.

Why the framework matters

OpenAI’s Preparedness Framework is among the company’s principal public explanations of how it intends to identify and manage risks from increasingly capable models. Robinson’s role in creating it gives his criticism particular relevance: he is challenging the wider industry after helping build one of the mechanisms meant to address that challenge.

Crypto Briefing reports that OpenAI is also facing criticism over its handling of advanced models and has paused the planned launch of GPT-6.1 Astra. The article does not establish that Robinson’s resignation caused the pause, but the two developments place renewed attention on whether the company’s existing safety processes are sufficient for its next generation of systems.

The immediate issue for OpenAI is credibility. A framework can reassure the public only if the company is seen as willing to follow it, give safety teams meaningful authority and act on warnings even when doing so slows a release. The recent organisational changes and staff exits have made that question more difficult for the company to answer.

From internal criticism to regulation

Robinson is not simply leaving the debate behind. He is reportedly working with Spitfire Strategies, a public-relations firm, on efforts to encourage stronger AI regulation and create outside pressure on laboratories to take safety more seriously.

That move points to a broader concern in his argument: voluntary safeguards may not be enough if companies face strong incentives to move ahead of competitors. If policymakers conclude that leading labs cannot reliably police themselves, external rules and oversight could become more prominent.

The next developments will help determine whether Robinson’s warning gains traction. OpenAI could respond publicly to his Atlantic essay, clarify how its safety teams now operate or provide an update on GPT-6.1 Astra. Further departures among safety staff would add to the impression of an unresolved conflict between rapid model development and independent risk management.