The most valuable people in artificial intelligence right now may not be the ones building the biggest models. They are the small number of researchers who claim to understand how to keep those models in check. And lately they have been changing teams. Two recent moves, in opposite directions between OpenAI and Anthropic, show just how fluid, and how contested, the market for safety expertise has become.
In one direction: Andrea Vallone, who built and led OpenAI's model-policy research group over a three-year tenure, has joined Anthropic's alignment team. Her work at OpenAI ran straight through the company's most sensitive territory, designing safety guardrails for GPT-4 and GPT-5, building reward systems for high-risk scenarios, and researching how models should respond to users in emotional distress. At Anthropic she is reported to be focusing on Claude's behaviour and mental-health-related safety, the very area where, as this week's cross-evaluation findings showed, every major model still stumbles.
In the other direction: OpenAI has hired Dylan Scandinaro, a former Anthropic safety researcher, as its new head of preparedness. Sam Altman announced the move himself, calling Scandinaro "by far the best candidate I have met, anywhere, for this role." The job reportedly pays up to $555,000 plus equity, a figure that tells you how the labs now value this work. Altman's framing was blunt: "Things are about to move quite fast and we will be working with extremely powerful models soon."
Taken together, the two hires are more interesting than either one alone. This is not a brain drain from one lab to another. It is a carousel. The same scarce pool of people who can credibly run a preparedness program or shape a model's policy behaviour keep circulating among the handful of labs that can afford them. When the supply of expertise is this thin, every senior hire is also a competitor's loss, and every departure raises quiet questions about what the person was unhappy with.
There is a notable thread in Vallone's move. She joins a lineage of OpenAI safety figures who left for Anthropic, reportedly reporting to Jan Leike, who himself departed OpenAI in 2024 amid public friction over how safety was being prioritised against shipping speed. Anthropic, founded by ex-OpenAI staff, has leaned into a "safety-first" identity, and it continues to attract people who want that to be the headline rather than the footnote. The pull is cultural as much as financial.
Yet the traffic flowing the other way complicates any tidy story about one lab being the safety conscience of the industry. OpenAI is paying top dollar and handing a prominent title to someone it poached from precisely that rival. Both companies are, in effect, bidding against each other for the credibility that safety researchers confer. The open question is whether this churn actually makes the resulting models safer, by spreading hard-won lessons across labs, or whether it simply means a tiny priesthood of experts keeps rewriting the same playbook under new logos while the models themselves race ahead.