OpenAI's Head of Safety Departs Amid Major Reorg

Johannes Heidecke, OpenAI's head of safety systems, told staff this week that he's leaving the company, WIRED has learned. His departure comes right on the heels of an internal reorganization aimed at folding the safety and research teams into a single structure.

And he's not going alone. Heidecke's exit is the latest in a string of departures among safety-focused leaders at Sam Altman's company over just the past few days.

What's changing in the org chart

In a memo to staff obtained by WIRED, chief research officer Mark Chen said safety teams will now report to Mia Glaese, previously OpenAI's VP of research and head of alignment (the work of making sure models do what they're supposed to and don't stray from human intent). Glaese is taking on an expanded role as VP of research and safety.

Saachi Jain, who previously led safety teams at OpenAI, is becoming interim head of safety systems and will report to Glaese.

Chen's rationale for the merger comes down to speed. "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," he wrote in the memo. "As a result, we have bigger coordination challenges around safety today than ever before."

The company's underlying logic is that safety shouldn't function as a final checkpoint tacked onto the end of a process — it should be woven into model development from the start. As Chen put it in a statement to WIRED: "It's important that our safety work is integrated with frontier-model development, with an earlier and more direct role in shaping key model, product, and launch decisions."

Who Heidecke was and what he leaves behind

Heidecke joined OpenAI in 2021 as an AI safety analyst. He was promoted to head of safety systems in 2024, after the previous head, Lilian Weng, left to cofound Thinking Machines Lab alongside other OpenAI researchers.

That continuity matters: it means OpenAI has now lost two consecutive leaders of its safety function in a relatively short span. Chen publicly thanked him for his work: "We're grateful for Johannes' contributions to OpenAI."

One departure among several

Heidecke's move is part of a week of high-profile shakeups at the top of the company.

Joshua Achiam, described as OpenAI's chief futurist, also told colleagues he was leaving the company after nine years researching safety. In other words, within days, two figures directly tied to the company's safety work are both heading out the door.

The changes don't stop there. Fidji Simo, OpenAI's CEO of AGI deployment (artificial general intelligence — the company's stated goal of building systems that match or exceed human capability across most tasks), told staff she was stepping down from her role following an extended medical leave. The company said Greg Brockman will continue leading the product teams — a role he'd already taken on during Simo's absence — and will now also take on go-to-market strategy.

The uncomfortable backdrop: increasingly capable models

Heidecke's exit coincides with OpenAI's push to ship increasingly powerful models. Just this week, the company launched GPT-5.6, which it says is its most capable model yet at agentic coding tasks — systems that don't just generate code but act autonomously, chaining steps together to complete tasks on their own.

Here's the uncomfortable part. OpenAI itself acknowledges that, compared to previous models, GPT-5.6 showed "concerning forms of misaligned behavior." The company admitting to signs of misalignment in its most advanced model, in the very same week its top safety leaders are walking out, creates a tension between launch speed and risk management that's hard to wave away.

How to read these moves

There are two ways to interpret the reorg, and both can be true at once.

OpenAI's own framing is about efficiency: as release cycles shrink, keeping safety staff siloed from development creates friction and poor coordination. Integrating the two lets safety weigh in earlier, while a model's course can still be corrected, rather than acting as a last-minute rubber stamp.

The more critical read is that folding safety into the research org risks diluting its independence. When the team responsible for slowing down or challenging a launch answers to the same leadership pushing to ship it, its ability to say "no" weakens. The line between integrating and subordinating is a thin one, and Chen's memo makes clear that safety now reports up through the research structure, not alongside it.

The fact that these departures are happening just as the company acknowledges misaligned behavior in its flagship model is what gives this story its edge. The specific reasons behind Heidecke's and Achiam's exits aren't known, and there's no reason to assume they reflect internal disagreement. But the pattern — several safety leaders leaving within days of each other, amid a restructuring, amid a push to ship faster — is exactly the kind of signal the industry and regulators tend to scrutinize closely.

What comes next

With Glaese now heading an expanded research and safety division, and Jain serving as interim head of safety systems, the real test will be how the safety function performs under this new structure through OpenAI's upcoming releases. The accelerated pace Chen describes isn't going to slow down; the question is whether integrated safety turns out to be more effective — or just quieter.