OpenAI's safety leaders are jumping ship

When the people hired to build the guardrails start jumping out of the moving car, you probably want to pay attention.

OpenAI launched its Superalignment team roughly a year ago with a sci-fi level mission: figure out how to control hypothetical super-intelligent AI before it turns on humanity. Now, the team's top two leaders, co-founder Ilya Sutskever and Jan Leike, have both walked out the door.

The receipt trail from November's coup

This isn't an isolated exit. The departures follow months of quiet fallout from the failed November 2023 boardroom coup led by Sutskever against CEO Sam Altman.

While Sutskever eventually apologized and joined 738 out of 770 employees on a public letter demanding Altman's return, an earlier copy obtained by The New York Times tells a different story. That version, containing 702 signatures, was missing several key names.

Among those missing were Superalignment researchers Jan Leike, Leopold Aschenbrenner, and William Saunders, along with renowned scientist Andrej Karpathy, Daniel Kokotajlo, and Cullen O'Keefe. Fast forward six months, and nearly every one of those holdouts has left the company. Leike confirmed on X that safety has taken a backseat at OpenAI, explaining why so many safety-focused minds are stepping away.

Safety has always broken OpenAI from the inside

Ideological clashes over safety are essentially woven into OpenAI's DNA. Back in 2021, Dario and Daniela Amodei took nine staffers and left to start rival Anthropic over similar disagreements.

November's board drama was fueled by the exact same rift. But instead of slowing down, Altman returned, and the nonprofit board members were swapped out for seasoned Silicon Valley operators.

OpenAI still insists it takes safety seriously, pointing to other staff members tasked with alignment. But when the people running your dedicated existential-risk unit walk away in lockstep, that narrative gets harder to sell.

Why it matters

For everyday developers and businesses using OpenAI's APIs, you will likely see faster product rollouts and fewer artificial throttles in the short term. However, the catch is that responsibility for frontier safety is quietly shifting onto your shoulders, meaning enterprise users will have to double-check high-autonomy agents for unaligned behavior themselves.

Across the industry, this brain drain cements Anthropic and independent safety labs as the go-to sanctuaries for cautious enterprise buyers. It draws a clear battle line between shipping at all costs and cautious deployment.

In our view, OpenAI's strategy makes complete sense if the goal is outrunning Google and keeping investors happy. But burning out or alienating your most vocal internal skeptics creates a dangerous blind spot: once you clear out everyone who is willing to pull the emergency brake, nobody is left inside to stop the train if things go sideways.