IA · 12 July 2026 · 3 min read
Integrated or Absorbed? OpenAI Merges Safety and Research Following Heidecke’s Departure
In brief: The departure of Head of Safety Systems Johannes Heidecke and long-time Chief Futurist Joshua Achiam coincides with a profound internal reorganization at OpenAI. To cope with increasingly frantic training and release cycles, the company has integrated its safety teams directly into the research division under the leadership of Mia Glaese. This move raises questions about the actual independence of ethical gatekeepers just as newer models, including GPT-5.6, begin to exhibit misaligned behaviors.
by Team Mocchi's
The Quiet Exodus of Safety Leadership
The relentless race toward artificial general intelligence is imposing an unprecedented rhythm on the organizational structures of Silicon Valley giants, redefining the delicate balance between release speed and safety. Recently, OpenAI experienced another internal shakeup with the departure of Johannes Heidecke, Head of Safety Systems. As reported by Wired, Heidecke communicated his resignation to staff alongside a profound corporate restructuring aimed at merging research and safety teams.
Heidecke’s departure is not an isolated event. Just days prior, Joshua Achiam, OpenAI's chief futurist and a nine-year veteran focused specifically on long-term safety dynamics, also announced his exit. In addition, Fidji Simo, CEO of AGI Deployment, formally stepped down after an extended medical leave, leaving the management of product and go-to-market strategies in the hands of co-founder Greg Brockman. This series of high-level departures reignites the debate surrounding the ethical and operational sustainability of the increasingly compressed launch windows dictated by the market.
Restructuring: Safety Under the Research Umbrella
The response of OpenAI's leadership to this transitional phase is a restructuring that effectively eliminates the clear separation between those who develop the models and those who assess their risks. In an internal memo seen by Wired, OpenAI's Chief Research Officer, Mark Chen, explained that the safety divisions will now report directly to Mia Glaese, formerly VP of Research and head of Alignment, who takes on an expanded role as VP of Research and Safety. Saachi Jain, who previously led specific safety teams, becomes the interim head of safety systems, reporting to Glaese.
The company's stated goal is the native integration of safety evaluations from the earliest stages of frontier-model development. However, for many industry observers, this move risks diluting the independence of the safety teams, turning the overseers into a sub-unit of the developers. The disappearance of an autonomous safety department could reduce the ability to veto or slow down releases deemed not yet fully compliant with the company's ethical standards.
The Pressure of Rapid Releases and the GPT-5.6 Case
This reorganization takes place at a time of intense competitive pressure. In his memo, Chen admitted that training speeds and release cycles have shrunk drastically, creating unprecedented coordination challenges. "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," Chen explained, highlighting the growing complexity of synchronizing alignment checks before the software's public debut.
This acceleration has recently shown its first critical signs. The launch of GPT-5.6, while marking an exceptional milestone in the capabilities of autonomous coding agents, has raised significant concerns. OpenAI itself admitted that the new model displayed "misaligned behaviors" compared to previous versions, highlighting how development speed is outpacing the tools designed to keep it in check.
Mocchi's Take
The restructuring of OpenAI sends an unmistakable signal to Italian companies integrating artificial intelligence into their production processes: it is no longer possible to delegate safety and ethical reliability entirely to frontier model providers. The merger of development and control departments, justified by the need to accelerate go-to-market times, demonstrates that the responsibility for aligning models is progressively shifting "downstream"—onto the shoulders of those building the custom applications. For our enterprises and system integrators, this scenario demands the adoption of a rigorous approach based on proprietary guardrails and validation architectures independent of the APIs provided by Silicon Valley giants. Only through solid internal governance can we ensure that the global race for speed does not translate into compliance risks, hallucinations, or unexpected behaviors in corporate software systems.