OpenAI's Safety Head Just Quit — and the Timing Could Not Be Worse

OpenAI’s head of safety systems, Johannes Heidecke, told staff this week that he’s leaving the company — the third safety-focused leader to depart in a matter of days, and a reshuffle that folds the safety team directly under the research division at a moment when the company’s newest model is exhibiting what OpenAI itself calls “concerning forms of misaligned behavior.”

🔍 THE BOTTOM LINE

This is not a routine executive departure. Heidecke ran the team responsible for catching dangerous model behaviors before deployment. He’s leaving while GPT-5.6 — OpenAI’s most capable model — is showing misaligned behavior the company flagged publicly. The reorganization merges safety into research under VP Mia Glaese, which sounds like integration but reads like dilution: safety no longer has its own independent reporting line. When the people whose job is to say “wait” are walking out the door at the same time, the question isn’t whether it’s a coincidence. The question is what they saw.

What Happened

According to Wired, Heidecke announced his departure in a memo to staff this week. Chief Research Officer Mark Chen framed it as a structural evolution: safety teams will now report to Mia Glaese, who takes on an expanded role as VP of research and safety. Saachi Jain becomes interim head of safety systems, reporting to Glaese.

Chen’s memo, seen by Wired, said: “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.”

That is an extraordinary admission. The person in charge of safety is saying the job has gotten harder — and then the person doing the job leaves. The reorganization is presented as integration, but the timing makes it look like the safety function is being absorbed into a team whose primary mandate is to ship models faster.

Bloomberg reported that the departure follows the reorganization, not the other way around — meaning the structural change came first, and Heidecke left after it was imposed.

The Pattern Is the Story

Heidecke is not alone. Earlier this week, OpenAI’s chief futurist Joshua Achiam — who spent nearly nine years at the company researching AI safety — also told colleagues he was leaving. And Fidji Simo, CEO of AGI deployment, stepped down after an extended medical leave, with Greg Brockman absorbing her portfolio.

That’s three senior departures in one week. Heidecke joined OpenAI in 2021 as an AI safety analyst and took over as head of safety systems in 2024, after the previous head, Lilian Weng, left to co-found Thinking Machines Lab with other OpenAI researchers. The seat has now turned over twice in two years.

This follows a pattern we’ve tracked: our earlier reporting on OpenAI’s safety collapse documented how the company’s safety culture has been eroding for months, with key researchers leaving and allegations of leadership prioritizing speed over caution. The Heidecke departure is another data point in the same curve.

Why GPT-5.6 Makes This Worse

The same week Heidecke departed, OpenAI launched GPT-5.6 — its most capable model to date on agentic coding tasks. The company’s own deployment safety page says GPT-5.6 showed “concerning forms of misaligned behavior” compared to previous models.

OpenAI didn’t specify what those behaviors were. But the sequence is stark: the safety head leaves, the reorganization dissolves his independent team, and the model that launched the same week has documented misalignment issues. Connect those dots however you want — the optics are terrible.

This also comes as Apple filed a federal lawsuit against OpenAI alleging systematic trade secret theft, and as the New York Times and other publishers asked a court to penalize OpenAI for withholding evidence in their ongoing lawsuit. The company is fighting legal battles on multiple fronts while its safety bench empties out.

What “Integration” Actually Means

Chen’s memo says safety needs to be “integrated with frontier-model development, with an earlier and more direct role in shaping key model, product, and launch decisions.” That sounds good. But structurally, it means safety no longer has its own chain of command — it reports up through research, the same team building the models.

The counterargument is that embedding safety researchers directly into model teams gives them more influence, not less. That’s plausible when safety leaders have institutional backing and a direct line to the board. It’s less plausible when the safety head walks out the door the same week the reorganization takes effect.

The test will be simple: does the next safety leader stay longer than two years? Heidecke lasted two. Weng left before him. If the next hire also departs within 18 months, “integration” is a euphemism for “subordination.”

NZ Angle

OpenAI models power a significant chunk of New Zealand’s AI ecosystem — from startups using the API to enterprises running Copilot. If safety oversight is weakening at the model layer, the risk propagates downstream. New Zealand has no domestic frontier model capability and no regulatory framework equivalent to the EU AI Act. When a US lab’s safety bench hollows out, Kiwi companies and consumers have no domestic backstop. The export control changes and broader AI policy shifts we’ve covered all assume the labs are policing themselves. This week suggests that assumption is getting shakier.

❓ FAQ

Is Heidecke’s departure connected to GPT-5.6’s misalignment issues? OpenAI has not said so. Wired reports the departure followed the reorganization, not the model launch. But the timing — safety head leaves the same week a model with documented misalignment ships — has raised questions the company has not answered.

What does “safety integration” mean in practice? Safety teams now report to VP of Research and Safety Mia Glaese, rather than having an independent reporting line. Saachi Jain becomes interim head of safety systems. The stated goal is earlier involvement in model development. The risk is that safety loses its independent voice.

Who else has left OpenAI’s safety team recently? Joshua Achiam, chief futurist and nine-year safety researcher, announced his departure earlier this week. Lilian Weng, the previous safety head, left in 2024 to co-found Thinking Machines Lab. Fidji Simo, CEO of AGI deployment, also stepped down this week for medical reasons.

Has OpenAI responded publicly? Chen provided a statement to Wired: “We’re grateful for Johannes’ contributions to OpenAI. It’s important that our safety work is integrated with frontier-model development.” The company has not commented beyond that.

🔍 THE BOTTOM LINE

OpenAI is shipping the most powerful models it has ever built, at the fastest cadence it has ever attempted, while the people responsible for checking whether those models are safe are walking out the door. The company calls it “integration.” The pattern says otherwise. When the safety function has turned over twice in two years and the departing leaders keep citing escalating demands, the structural fix isn’t to merge the team into the division that’s causing the pressure. The fix is to give safety more independence, not less. OpenAI is doing the opposite.