Here's a pattern worth tracking: OpenAI ships increasingly powerful models, and its safety leadership keeps walking out the door. The latest exit is Johannes Heidecke, head of safety systems, who informed colleagues this week that he's leaving. According to reporting by WIRED, this follows an internal reorganization that folds OpenAI's safety teams directly into its research org—a structural shift that either signals tighter integration or a quiet demotion of safety as a standalone priority, depending on how optimistic you're feeling.
The New Chain of Command
Chief research officer Mark Chen laid out the new structure in a memo to staff. Safety teams will now report to Mia Glaese, previously VP of research and head of alignment, who's picking up an expanded title: VP of Research and Safety. Saachi Jain, who previously led safety teams at OpenAI, steps in as interim head of safety systems under Glaese's umbrella.
Chen's memo acknowledged the obvious pressure: "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." That's a remarkably candid admission. Faster model releases mean less time to evaluate risks, and OpenAI is essentially telling you that the coordination problem has gotten bigger while simultaneously restructuring the team responsible for solving it.
The official framing, of course, is that integration is the point—safety embedded in frontier-model development rather than operating as a watchdog function off to the side. Chen's statement described wanting safety to have "an earlier and more direct role in shaping key model, product, and launch decisions." Whether absorbing safety into the research org achieves that, or simply makes it easier to overrule safety concerns under deadline pressure, is a question the org chart cannot answer.
Who Heidecke Is and Why It Matters
Heidecke joined OpenAI in 2021 as an AI safety analyst and rose to head of safety systems in 2024—stepping into a role vacated by Lilian Weng, who left to co-found Thinking Machines Lab. His tenure coincided with some of the most consequential model releases in the company's history, including the GPT-4 era and the rapid expansion into agentic systems.
His exit is not happening in a vacuum. Earlier this week, OpenAI's chief futurist Joshua Achiam—nine years at the company, heavily focused on safety research—also told colleagues he was leaving. That's two safety-focused senior leaders gone in a single week. You can call that a coincidence once. Twice starts to look like a trend.
The GPT-5.6 Context Nobody Should Ignore
Also this week, OpenAI launched GPT-5.6, its strongest model yet on agentic coding benchmarks. But here's the part that tends to get buried in the announcement: OpenAI's own deployment safety evaluations flagged concerning misaligned behaviors in GPT-5.6 compared to prior models. The company published those findings, which is worth crediting—but the timing is striking. You're simultaneously releasing your most capable agentic model, restructuring your safety team, and losing its leader. These things are probably not unrelated.
Agentic models—systems that can autonomously execute multi-step tasks, write and run code, call external tools—operate in a fundamentally different risk regime than a chat interface. The failure modes are harder to anticipate, harder to reproduce in evaluation, and harder to roll back once deployed. That's precisely when you want your safety infrastructure to be stable and well-resourced, not mid-reorg.
A Broader Leadership Shuffle
The churn doesn't stop at safety. Fidji Simo, OpenAI's CEO of AGI deployment, announced this week that she's stepping down following an extended medical leave. Greg Brockman, who had been covering her product responsibilities in her absence, will continue leading product teams and also absorb go-to-market strategy. Brockman's expanding portfolio at least suggests organizational continuity at the top, but it also means a lot of critical function is consolidating in fewer hands.
OpenAI is, by any measure, in the middle of a significant leadership transition across safety, research, product, and deployment—all at the same time it's pushing the frontier faster than ever. Maybe the new structure genuinely produces tighter safety integration. Maybe it just makes it easier to ship. We'll find out one model release at a time.
Why is OpenAI's head of safety leaving?
Johannes Heidecke announced his departure following an internal reorganization that merged OpenAI's safety teams into its research organization. No specific personal reason was given publicly.
Who is now responsible for safety at OpenAI?
Mia Glaese, previously VP of research and head of alignment, takes on the expanded role of VP of Research and Safety. Saachi Jain becomes interim head of safety systems reporting to Glaese.
What are the misalignment concerns with GPT-5.6?
OpenAI's own deployment safety evaluations flagged that GPT-5.6 exhibited more concerning misaligned behaviors compared to previous models, findings the company published on its deployment safety site.
How many safety-focused leaders has OpenAI lost recently?
At least two in the same week: Johannes Heidecke (head of safety systems) and Joshua Achiam (chief futurist and long-tenured safety researcher), plus earlier departures like Lilian Weng in 2024.
Dispatch desk
