OpenAI’s head of safety systems, Johannes Heidecke, is departing the company following an internal restructuring that merges the safety and research teams under a single leader, according to a report published on Friday. The move represents yet another shift in how the artificial intelligence lab approaches safety oversight, and it continues a pattern of senior safety figures leaving or being reorganized out of the company over the past two years.
Chief Research Officer Mark Chen informed staff in a memo that safety teams would now report to Mia Glaese, whose title has been expanded to VP of Research and Safety, a newly created position that signals the company's intent to keep safety as a named priority even under the new structure. Saachi Jain has been named interim head of safety systems while the company searches for Heidecke’s permanent replacement.
The role and tenure of Johannes Heidecke
Heidecke joined OpenAI in 2021 as an AI safety analyst and worked his way up to become head of safety systems in 2024, succeeding Lilian Weng. During his tenure, he was responsible for overseeing model alignment — the process of ensuring AI systems behave in accordance with human intentions — as well as rule-based reward systems, which provide explicit guidelines for model behavior. He also led the company’s preparedness evaluations, which assess potentially dangerous model capabilities before deployment.
In his memo, Chen thanked Heidecke for his contributions and emphasized that it is “important that our safety work is integrated with frontier-model development, with an earlier and more direct role in shaping key model, product and launch decisions.” This integration philosophy forms the backbone of the current restructuring, but critics argue that it may reduce the independence of safety oversight.
A pattern of disbandments and departures
Heidecke’s exit is not an isolated event. Over the last two years, OpenAI has repeatedly reorganized or dissolved its safety-focused teams. In 2023, the company announced the Superalignment team with a high-profile pledge to allocate 20% of its compute resources to safety research. However, the team was dissolved in May 2024 after its co-leads, Ilya Sutskever and Jan Leike, departed. Leike publicly stated that “safety culture and processes have taken a backseat to shiny products,” a critique that resonated with observers who track the company’s shifting priorities.
Following that dissolution, the AGI Readiness team was formed, but it lasted only a few months. Its leader, Miles Brundage, resigned in October 2024. The Mission Alignment team, which succeeded the Superalignment efforts, was disbanded in February 2026 after a mere 16 months. Its leader, Joshua Achiam, was given a new title: “chief futurist.” The cumulative effect of these changes has been a significant turnover in safety leadership and a perceived weakening of institutional safeguards.
In April of this year, OpenAI lost its product chief, the head of Sora (its video generation tool), and its enterprise CTO on a single day. More recently, Fidji Simo, OpenAI’s chief of applications, stepped down this month citing a prolonged medical recovery. These departures, combined with the safety team restructuring, paint a picture of an organization in flux.
The case for integration versus independence
Mark Chen’s stated logic for merging safety into research is that embedding safety teams inside research gives them a seat at the table from the earliest stages of model development, rather than treating safety as a final checkpoint before launch. Proponents of this approach argue that it leads to more practical safety interventions that are better aligned with engineering realities.
However, critics counter that a safety team reporting inside a research hierarchy has less structural independence and less leverage to delay or block a product release if safety concerns arise. In previous structures where safety reported separately — either to the CEO or to a dedicated safety committee — the team could raise red flags without fear of being overruled by research priorities. The new arrangement blurs those lines and may reduce accountability.
This tension is not unique to OpenAI. Across the AI industry, companies are grappling with how to organize safety functions. Some, like Anthropic, have maintained separate safety teams with independent reporting lines. Others, like Google DeepMind, have integrated safety into research while also maintaining external oversight. The debate reflects deeper questions about whether AI safety is best served by integration or independence.
External pressures intensify
Heidecke’s departure arrives as OpenAI faces growing external scrutiny. Forty-two state attorneys general have opened an investigation into the company, serving a subpoena related to advertising, user data, and internal policies. This investigation comes shortly after OpenAI confidentially filed for a stock market listing, a move that has raised questions about the company’s governance and transparency.
The legal pressure adds to a broader regulatory environment that is increasingly focused on AI safety. In the European Union, the AI Act is set to impose strict requirements on high-risk AI systems, including general-purpose AI models like those developed by OpenAI. Compliance will require robust safety processes, making the timing of the safety team restructuring particularly sensitive.
In response to these challenges, OpenAI launched a Safety Fellowship on April 6, inviting external researchers to conduct independent safety and alignment work at the lab. The program is intended to bring fresh perspectives into the company’s safety efforts, but it remains to be seen whether it can compensate for the loss of internal advocates like Heidecke.
Leadership churn continues
The safety systems role has now seen multiple leaders in a short period. Lilian Weng, who held the position before Heidecke, left OpenAI to join Thinking Machines Labs, an AI startup founded by former OpenAI CTO Mira Murati. Murati herself has been vocal about governance issues, warning publicly that AI governance is lagging behind model capability. Her comments echo the concerns of many former OpenAI employees who feel that the company’s focus has shifted too heavily toward commercialization at the expense of safety.
Meanwhile, the company’s internal culture has come under scrutiny. Reports have described a high-pressure environment where conflicts between research and safety teams are common. The latest restructuring may aim to resolve these conflicts by aligning incentives, but it also risks suppressing dissenting voices.
Looking ahead, OpenAI’s approach to safety will continue to be a bellwether for the industry. As models become more powerful and applications more widespread, the decisions made today about organizational structure could have long-lasting consequences for the responsible development of artificial general intelligence. For now, the departure of another senior safety leader raises questions that the company has yet to fully answer.