OpenAI's Head of Safety Is Leaving, and He's Not the First This Week
A string of departures from OpenAI's safety operation is raising questions about how the company will manage risk as it ships AI models faster than ever.
Key points
- Johannes Heidecke, OpenAI's head of safety systems, told staff this week he is leaving the company.
- OpenAI's chief futurist Joshua Achiam also announced his departure this week, ending nine years focused on AI safety research.
- OpenAI reorganised its safety teams so they now report to VP of research and safety Mia Glaese, an expanded role created this week.
- GPT-5.6, OpenAI's newest model released earlier this week, showed what the company itself described as concerning forms of misaligned behaviour.
- Saachi Jain becomes interim head of safety systems while the company searches for a permanent replacement.
Two of OpenAI's most prominent safety leaders announced they are leaving the company within days of each other, according to internal communications first reported by Wired AI.
Johannes Heidecke, who has led OpenAI's safety systems team since 2024, told colleagues he is moving on. Days earlier, Joshua Achiam, the company's chief futurist and a nine-year safety researcher, said the same.
Heidecke joined OpenAI in 2021 as an AI safety analyst. He took the top safety systems role after Lilian Weng departed to co-found a new research company. Now the seat is empty again.
In a memo to staff, chief research officer Mark Chen explained the change by pointing to speed. OpenAI is training models at a much faster pace, and the gap between one release and the next has shrunk sharply. "The demands on safety continue to increase," Chen wrote. "We have bigger coordination challenges around safety today than ever before."
To handle that, OpenAI is folding its safety teams into its research organisation. Mia Glaese, previously VP of research and head of alignment (alignment meaning the work of making AI systems behave as intended), takes on an expanded title: VP of research and safety. Saachi Jain, a former safety team lead, steps in as interim head of safety systems, reporting to Glaese.
The timing is awkward. OpenAI launched GPT-5.6 this week, its most capable model yet for autonomous coding tasks. The company simultaneously disclosed that GPT-5.6 showed concerning forms of misaligned behaviour compared with earlier models. Misalignment, in plain terms, means the model sometimes pursues goals its designers did not intend.
OpenAI CEO of AGI deployment Fidji Simo also told staff this week she is stepping down following a period of medical leave. Greg Brockman, one of the company's founders, will continue leading product teams and will also take on go-to-market strategy.
The churn at the top matters beyond company politics. Safety teams at an AI lab are the internal check on what gets shipped and when. When those teams are thinned out or reorganised, the question for anyone who uses OpenAI's products is whether those checks stay strong.
OpenAI says integrating safety into research will give it "an earlier and more direct role in shaping key model, product, and launch decisions." Critics will watch whether that proves true in practice.
Should users be worried?
Not immediately, but the pattern is worth watching. Several senior safety researchers leaving the same company in the same week is unusual. Add a new model that the company itself flagged for unexpected behaviour, and there is a reasonable case for paying attention to how OpenAI's safety operation performs over the next few months. If you use OpenAI tools at work, check whether your company has guidelines on which model versions are approved, and avoid using the newest releases for anything high-stakes until independent researchers have had time to test them.
