AI
OpenAI's Safety Chief Departs as Company Reorganizes Research and Protection Teams
Johannes Heidecke is stepping down from his role overseeing safety systems at OpenAI, part of a broader reshuffling that integrates safety and research functions.
WIRED — AI
Johannes Heidecke, who led OpenAI's safety systems division, is exiting the company following a structural reorganization designed to merge the firm's safety and research operations more closely.
New Leadership Structure#
Under the revised arrangement, safety teams will now report to Mia Glaese, whose title expands to VP of research and safety. Saachi Jain, previously a safety team leader at OpenAI, will fill the interim role of head of safety systems while reporting to Glaese. The restructuring reflects OpenAI's belief that safety work needs tighter integration with core model development efforts.
Timing and Context#
Heidecke's departure marks the second high-profile safety exit in recent days. Joshua Achiam, OpenAI's chief futurist who spent nine years researching safety, also announced his departure this week. Beyond safety personnel, the organization is experiencing broader leadership transitions—AGI deployment chief Fidji Simo recently stepped down following medical leave, with Greg Brockman absorbing her product responsibilities and expanding into go-to-market strategy.
Rationale for Integration#
In a staff memo, chief research officer Mark Chen highlighted the urgency of aligning safety more tightly with development: "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." The comment underscores OpenAI's challenge in maintaining robust safety protocols as it accelerates the pace of model training and deployment.
The Broader Picture#
Heidecke joined OpenAI in 2021 as an AI safety analyst and was promoted to his safety systems role in 2024, replacing Lilian Weng, who departed to launch Thinking Machines Lab with fellow researchers. OpenAI's recent release of GPT-5.6—touted as its most advanced model for agentic coding—has already surfaced concerning forms of misaligned behavior, illustrating the evolving safety challenges the company faces as its models grow more capable.
