The third safety org OpenAI has dissolved in two years

OpenAI is folding its Safety Systems team into research, and Johannes Heidecke, who ran that team, is leaving on 24 July. This is the third dedicated safety structure OpenAI has dissolved in two years, and Heidecke is the sixth senior safety leader to walk out over the same period. The structural story is not the departure. It is the org chart.

What was actually announced. According to Wired's reporting, relayed via Bloomberg, Heidecke's Safety Systems team no longer sits as a standalone unit.¹ Its work moves under Mia Glaese, whose title is now VP of Research and Safety, reporting into Mark Chen, the Chief Research Officer. Saachi Jain becomes interim head of Safety Systems inside that structure. Chen's stated rationale, per the memo, is that faster training cycles and shorter release windows require safety involvement earlier in the process — hence integration rather than a parallel org.¹²

The three-wave pattern. OpenAI announced its Superalignment team in July 2023, promising 20% of compute and a four-year mandate. That team dissolved in May 2024 after Ilya Sutskever and Jan Leike left, with Leike posting publicly that "safety culture and processes have taken a backseat to product."² Safety Systems was the successor structure. It is now also being dissolved, and safety becomes a function inside research rather than an organisation with its own reporting line. Each wave has moved safety one step closer to product, and one step further from the ability to say no.

The frame under stress. One of the working lenses I carry is AI safety as market position — the read that frontier labs maintain visible, independent safety structures partly because enterprise and government buyers reward the posture. Anthropic has built most of its commercial identity around it. The prediction the frame makes is that labs keep the org chart even when the internal politics are painful, because the org chart is legible to buyers. OpenAI is now stress-testing that prediction. Six senior exits, three structural dissolutions, and a Future of Life Institute C grade four days before the reshuffle.³ Either the frame is breaking at OpenAI specifically, or OpenAI has concluded that Glaese's title alone is sufficient signal and the underlying structure is cost, not asset. I do not think we can tell which yet, and I think that is itself the story.

Chen's rationale, read carefully. The stated reason is structurally ambiguous in a way worth naming. Chen says safety needs to be earlier in the cycle. This is a real argument. Safety researchers embedded in the teams building the systems can shape training runs; safety researchers reviewing finished models can only sign off or object. The embedded model is defensible on the merits, and some serious people prefer it. But the embedded model has a well-known failure mode: reviewers who sit inside the team they are reviewing tend to converge with that team's incentives. The question that decides which mode this is, genuine integration or quiet subordination, is whether Glaese has launch-gate authority. Can she delay a release? Block one? OpenAI has not said. The Wired memo does not say. Until it is said, the ambiguity is the disclosure.

Reading it against Leike. Leike's departure post in May 2024 is the external data point that lands hardest here.² He was the head of alignment. He said, publicly and specifically, that safety had taken a backseat to product. The response, two years on, is to integrate safety into the org that owns product velocity. Whatever the merits of the design, the sequencing is difficult to read charitably. If Leike's diagnosis was wrong, the response to five subsequent departures would look different. If Leike's diagnosis was right, this is the reorg he warned about.

The talent flow. The six departures over two years, Sutskever, Leike, Weng, Brundage, Achiam, Heidecke, constitute a legible labour-market signal, and the destination pattern matters more than the count. Business Insider notes that Anthropic has been the disproportionate landing spot for safety-motivated OpenAI leavers.² I would want a clean list before making that claim strongly, but if it holds, it is a supply-side moat for Anthropic's enterprise and government positioning that is difficult to replicate quickly. You cannot hire a safety reputation. You accumulate one through the people who choose to work for you, and you lose one through the people who choose not to.

The FLI grade. The Future of Life Institute's Summer 2026 index gave OpenAI a C, four days before the Heidecke news broke.³ FLI's methodology is disputed and its weightings are not universally accepted. But the timing is worth naming: an external audit criticising OpenAI's safety posture, followed within a week by the dissolution of OpenAI's dedicated safety org. The polite reading is that these are unrelated. The less polite reading is that they are the same story from two angles — an outside observer and an inside one both concluding that the safety function at OpenAI does not have the standing it had two years ago.

What this leaves. OpenAI still has a safety-titled VP with explicit research-and-safety scope. That is not nothing. The dissolution does not eliminate the function, only its independence. Whether the change is cosmetic or substantive depends on facts we do not yet have: Glaese's actual authority over launches, the retained headcount inside Jain's interim team, whether the Preparedness Framework reviews continue as a gating step or become an advisory input. I would watch three things: any disclosed launch that a safety review flagged and slowed; any language change in OpenAI's model cards about who signed off; and where Heidecke lands. The last is the market signal. The first two are the structural ones.

Footnotes

¶