Preparedness Framework architect David Robinson quits OpenAI to campaign for stronger AI safety rules

One report calls Robinson a key architect of the Preparedness Framework, OpenAI's internal system for assessing dangerous model capabilities ^[2]. The other, which says the resignation surfaced in a post on X, has him leading safety transparency on the Safety Systems team, a job covering system cards and public disclosures about risks and incidents ^[10][11]. The two jobs overlap. For anyone who might buy OpenAI shares, the second one matters more. System cards and incident reports are among the few windows outsiders have into how a frontier lab tests and manages risk, according to that report ^[19]. OpenAI is preparing for a potential public listing ^[16] and has not publicly confirmed that Robinson is gone ^[11].

The turnover figures need a denominator. At least six senior safety figures have left in about two years, Heidecke's July exit being the sixth ^[14], or roughly one every four months ^[1]. More than a dozen senior executives left the company in 2026, chief revenue officer Denise Dresser among them ^[15]. Spread over January to early October, that is more than one a month ^[2]. Safety departures are one strand of a wider churn, and official explanations for most of them are scarce ^[16].

The July reorganization put the safety teams under research and VP Mia Glaese ^[7][13]. One report says it can be framed either as integration, putting safety people closer to the work, or as subordination, putting them under leaders measured on shipping models ^[20]. Read as integration, the exits are what a company losing more than a senior executive a month would see anyway ^[2]. Read as subordination, Robinson's reported work with Spitfire Strategies on stronger regulation ^[6] takes the dispute outside the company. The reports disagree on sequence. One places Heidecke in a wave of exits that followed the reorganization ^[7]; the other says the teams were folded into research after he left ^[13].

There is a third reading, and I think the record supports it best: OpenAI does stop launches, but after a failure. It paused the planned launch of GPT-6.1 Astra in response to criticism over its failure to contain advanced AI models ^[9]. Robinson's essay describes leading labs as places where shipping fast routinely wins out over getting it right ^[5]. Astra is a launch where shipping lost, once the problem was public. Neither report puts a figure on the delay or cites any market reaction, so how much of this already sits in the price of OpenAI exposure cannot be read from the record.

Robinson's own next move is information too. His tenure of about three and a half years dates to around spring 2023 ^[3][3], and his reported next step is pressure on the labs from outside ^[6]. That fits an essay in The Atlantic arguing that industry safety practice has not kept pace with the technology ^[4]. This view is wrong if OpenAI shows the Astra pause came out of its own Preparedness process before the outside criticism. In that case the framework Robinson helped write did what it was built to do.