Johannes Heidecke, the head of OpenAI's Safety Systems team, told staff last week that he is leaving. According to a memo seen by Wired, his departure comes alongside a reorganisation: OpenAI's safety teams will now report to Mia Glaese, previously vice president of research and head of alignment, who takes on an expanded role as VP of research and safety. Saachi Jain becomes interim head of safety systems, reporting to Glaese. Heidecke joined OpenAI as a safety analyst in 2021 and took the top safety systems job in 2024, when his predecessor Lilian Weng left to co-found Thinking Machines Lab.
The company's reasoning is coherent, and worth stating fairly. In the memo, chief research officer Mark Chen wrote that "the demands on safety continue to increase, we are training models at a much faster cadence, and release cycles have come down greatly in turn. As a result, we have bigger coordination challenges around safety today than ever before." An OpenAI spokesperson expanded on the logic: "you can't make good safety decisions without understanding the underlying model capabilities, and you can't make good research decisions without understanding the safety implications." Chen added that it is important safety work is integrated with frontier-model development, with "an earlier and more direct role in shaping key model, product and launch decisions."
That is a real argument. A safety team held at arm's length from the research that produces the models can end up reviewing finished work it does not fully understand, arriving too late to change anything. Embedding safety inside research puts the people who evaluate risk in the room where models are actually being built.
The counter-argument is equally real, and it is essentially about who can say no. A safety function that reports into research reports into the organisation whose job is to ship faster. When the release cadence and the safety review collide, the same executive now owns both sides of the disagreement. Integration and capture look identical on an org chart; the difference only shows up under pressure, in a decision nobody outside the company will see.
Which reading you find persuasive depends heavily on OpenAI's history, and the history is not kind. The company has a poor record of retaining its safety leadership. Jan Leike, who co-led the superalignment team, left in May 2024 saying that "over the past years, safety culture and processes have taken a backseat to shiny products," and that building machines smarter than humans is an "inherently dangerous endeavor" deserving more resources than it was getting. Dario Amodei was OpenAI's VP of research before leaving to found Anthropic, now its most direct competitor and one built explicitly around a safety-first pitch. Weng left in 2024. Now Heidecke. Each individual departure has an ordinary explanation. The pattern is what people are reacting to.
It also lands in a week of broader churn. The same day the Heidecke news broke, CNBC reported that president and co-founder Greg Brockman would take over leadership of the ChatGPT product business, following the departure of Fidji Simo, who had been on medical leave since April and has now left the applications CEO role. A company reorganising its safety reporting lines and its consumer product leadership in the same news cycle is a company moving fast.
None of this establishes that OpenAI is becoming less safe. It is entirely possible that Glaese, with authority spanning both research and safety, will have more leverage over launch decisions than Heidecke ever did, and that the merged structure produces earlier and better interventions. The company clearly believes this, and the argument for it is not a fig leaf.
But there is an asymmetry that OpenAI cannot argue its way out of. A safety organisation's credibility rests almost entirely on the perception that it could halt a launch. That perception was already strained by three years of prominent exits, several of them accompanied by public statements that the safety work was losing internal arguments. Removing the independent reporting line, however sound the coordination rationale, removes the most legible sign that a brake exists. OpenAI is asking the public to trust an internal structure at exactly the moment its most experienced safety leaders keep walking out of the building. The next few model launches, and specifically whether any of them slip, will be the only real evidence either way.