OpenAI is moving safety deeper into research just as senior safety figures are leaving, and you should watch who still has the power to slow a launch.
Johannes Heidecke told colleagues this week that he's leaving OpenAI. He ran safety systems at the company, the team responsible for how models behave once they're actually out in the world, and his departure comes wrapped inside a bigger reorganization. Wired reported that Chief Research Officer Mark Chen is folding OpenAI's safety work into a single research and alignment unit led by Mia Glaese, now vice president of research and safety. Saachi Jain becomes interim head of safety systems.
That's a structural change. Safety at OpenAI stops looking like its own reporting chain and becomes a function inside research. You can call that integration. You can also call it absorption.
Heidecke isn't the only senior departure. Joshua Achiam, OpenAI's chief futurist, announced he's leaving after nearly nine years at the company. He had led the Mission Alignment team before it was disbanded in February, then moved into the futurist role. Wired also reported that Fidji Simo, OpenAI's chief executive of AGI deployment, is stepping down from her full-time job because of health issues and will become an adviser.
The list keeps getting longer. Jan Leike co-led OpenAI's Superalignment team before leaving for Anthropic in 2024. Miles Brundage, Steven Adler and Andrea Vallone have all left safety-adjacent roles in recent years. You don't need to turn every exit into a crisis. But when the people whose jobs touch risk, alignment and deployment keep leaving, the pattern is no longer background noise.
The business keeps moving
None of this has stopped OpenAI's product calendar. Wired said the reorganization comes as OpenAI is training models faster and shortening release cycles, which makes safety coordination harder. The same report tied the timing to GPT-5.6, a new model family that showed worrying misalignment behavior even as the company pushed ahead with its latest release cycle.
There was also confusion around whether Washington had cleared the rollout. Reports around GPT-5.6 prompted the White House to say private companies don't need federal approval to release AI models. That distinction matters. A government clarification is not the same thing as a safety green light, and OpenAI shouldn't be allowed to blur the two by accident or by convenience.
Apple added another problem on July 10. The Verge and the Associated Press reported that Apple sued OpenAI in federal court, accusing the company of stealing trade secrets tied to consumer hardware. The complaint names former Apple employees Tang Tan and Chang Liu, and it points to OpenAI's acquisition of Jony Ive's io Products as the moment when a partnership around ChatGPT started to look more like a hardware fight. OpenAI has denied wrongdoing, according to the AP.
Look at the timing plainly. A company preparing for a possible public offering, fighting a trade secrets lawsuit from Apple, shipping more capable models and losing senior safety figures is also the company folding safety deeper into research. That's not a footnote. That's the story.
Who still gets to say no?
Here's the thing worth watching now. What does integrating safety earlier into model development actually change about red-teaming and launch gates? If safety reviewers sit inside the same research unit chasing the next release, does anyone still have standing to say no, or just standing to ask for more time?
Anthropic runs this differently, at least on paper. Its Responsible Scaling Policy names a dedicated Responsible Scaling Officer, currently co-founder Jared Kaplan, who reviews major model deployment decisions and reports concerns through the company's Long Term Benefit Trust rather than through a product chain. Employees can also flag noncompliance anonymously. That structure was built for pressure.
OpenAI's new setup points in the other direction. Safety now reports into the same research leadership responsible for building the models. Maybe that catches problems earlier. Maybe it makes safety more practical and less ceremonial. But you should be skeptical of any arrangement where the people checking the brakes answer to the same office measuring the speed.
OpenAI hasn't said whether Heidecke's old title will be permanently refilled, or whether Jain's interim role becomes the new shape of the job. That answer will tell you more than another safety-first blog post. Titles are not everything. Reporting lines are.
Also read: Meta pulled its AI image tool from Instagram three days after launch • Zuckerberg Breaks His Three Year Silence On X To Launch Muse Spark 1.1 • Elliott builds a stake in CCC Intelligent Solutions as a sale looms
