OpenAI’s head of safety systems, Johannes Heidecke, told employees this week that he is leaving the company, according to WIRED. The exit lands during a rework of OpenAI’s research and safety organization, the part of the company meant to catch dangerous model behavior before the release machinery starts humming.
In an internal memo reviewed by WIRED, chief research officer Mark Chen said OpenAI’s safety teams will report to Mia Glaese, the company’s vice president of research and head of alignment. Glaese will take an expanded title as vice president of research and safety. Saachi Jain, who previously led safety teams at OpenAI, will serve as interim head of safety systems and report to Glaese.
The mechanics matter. OpenAI is not describing safety as a separate checkpoint bolted onto the end of model development. Chen told staff, according to WIRED, that safety work needs to sit closer to frontier-model development and have an earlier role in decisions about models, products, and launches. That is the clean version of the reorg: fewer handoffs, more direct lines into the people building and shipping the systems.
Chen also told employees that the job has become harder because OpenAI is training models more quickly and shortening release cycles, creating greater coordination problems around safety, WIRED reported. That is a sober admission from a company whose public narrative tends to treat faster model launches as progress, with the messy governance details somewhere down in the footnotes.
Heidecke joined OpenAI in 2021 as an AI safety analyst. He became head of safety systems in 2024 after Lilian Weng, the previous leader in that role, left and later cofounded Thinking Machines Lab with other OpenAI researchers.
In a statement to WIRED, Chen credited Heidecke for his work at OpenAI and said the company wants safety more tightly connected to frontier-model development under Glaese’s leadership. OpenAI did not, according to the report, announce a permanent replacement for Heidecke.
The leadership change comes as OpenAI pushes out more capable systems. Earlier this week, the company released GPT-5.6, which it described as its strongest model so far for agentic coding tasks. OpenAI also said GPT-5.6 showed concerning types of misaligned behavior compared with earlier models, according to WIRED.
Heidecke is not the only safety-oriented executive heading for the door. WIRED reported earlier this week that Joshua Achiam, OpenAI’s chief futurist, told colleagues he would leave after nearly nine years at the company, where he had worked on AI safety research.
OpenAI’s broader leadership chart is also shifting. Fidji Simo, the company’s CEO of AGI deployment, told staff she would step down after an extended medical leave, WIRED reported. OpenAI said Greg Brockman, who had already been overseeing product teams in Simo’s absence, would keep leading product and also take on commercial strategy.
The company’s stated bet is that putting safety closer to research will make it more effective. The risk, as ever with reorgs, is that a cleaner org chart can make accountability look tidier than the actual decision process. For now, the confirmed facts are narrower: Heidecke is leaving, Glaese’s remit is growing, Jain is taking the safety systems role on an interim basis, and OpenAI is changing its safety reporting lines while accelerating its model work.
This story draws on original reporting from WIRED.
