Johannes Heidecke, OpenAI’s head of safety, has left the company in the wake of an internal reorganization, people familiar with the matter said, the second senior leadership change at the AI firm within a week of its release of GPT-5.6. His exit follows a restructuring that formally folded safety oversight into the research department, according to WIRED, which first reported the shift, and Bloomberg, which confirmed it.
OpenAI will no longer maintain a standalone safety organization, the reports said. The safety function now sits inside the research group, reporting through the same chain of command that builds and ships the company’s models. The change ends a structure OpenAI had maintained for years, in which a dedicated team reviewed models before release and served as an independent check on the researchers who trained them.
Heidecke, who led OpenAI’s safety work, had been a public face of the company’s commitments to evaluating models before deployment. He joined the company’s safety efforts as they grew from a small team into one of the largest in the industry, people familiar with his work said, and he helped set the internal processes for testing and red-teaming new systems. His team’s remit had expanded in recent quarters to cover not just model behavior but also the security of the infrastructure behind the company’s consumer products, a portfolio that put him at the center of some of the company’s most sensitive internal debates.
The reorganization is the latest step in a long drift away from OpenAI’s original governance model. The company was founded as a nonprofit with a charter pledging to develop AI safely and for the benefit of humanity, but it has repeatedly restructured its safety apparatus as commercial pressures have grown. In 2024, the superalignment team that worked on long-term risks was disbanded, and its leaders, including co-founder Ilya Sutskever and researcher Jan Leike, left the company. Leike said at the time that safety culture had taken a back seat to product velocity. Each departure since then has been read as another marker of the same trade.
Supporters of the new structure argue that embedding safety inside research speeds iteration and keeps the people who understand the models in charge of their risks. Evaluations are most useful when they run continuously through the training process, they point out, not as a final gate held by a separate team. “Safety is not being de-emphasized; it is being reorganized,” said a person close to the company. Skeptics say the change removes an independent check and puts the evaluator in the pocket of the builder, a conflict that regulators in Europe and the United States have been watching closely as they draft rules for frontier AI.
The timing has drawn attention inside the company. GPT-5.6 shipped less than a week before Heidecke’s exit, and the departure of two senior executives in that window has raised questions among staff about the stability of the leadership team, according to people familiar with internal discussions. The model itself is rolling out to consumers and business customers on schedule, OpenAI has said, and the company has not indicated any change to its release plans. OpenAI did not comment on the personnel moves beyond confirming the reorganization to WIRED.
The industry has been moving in the same direction, though at different speeds. Anthropic maintains a separate safety team and has positioned its governance structure as a selling point. Google DeepMind has integrated safety review into its development pipeline. Microsoft has set up its own AI safety office. Analysts said OpenAI’s move aligns it with the industry’s dominant approach while exposing it to criticism that it is lowering its own bar. The distinction matters to enterprise buyers, some of whom have told procurement teams they want to understand how safety is governed before committing to a model.
The reorganization also shifts how OpenAI talks about safety in public. The company said it remains committed to rigorous evaluation and has pointed to its external testing programs, including work with outside researchers, as evidence that scrutiny continues without a standalone team. Whether that message holds depends on the details of the new structure: who signs off on releasing a model, what happens when an evaluation finds a serious problem, and how much authority the researchers who own safety now have over the schedule. None of those questions has been answered publicly, and the answers will determine whether the change is a realignment or a rollback.
What remains to be seen is how the change affects GPT-5.6 and the models that follow. Regulators in the European Union are expected to ask how safety assurances will be audited without a dedicated team, and researchers who track the company will watch whether more safety staff follow Heidecke out the door. For a company that built part of its brand on the promise of careful AI governance, the test is whether safety can survive inside the engine room.
