Head of Model Safety.

The lab had reached the point where safety could no longer sit as a working group inside research. Investors, the board, and the lab's external commitments all required a single senior owner accountable for the model-safety surface — evaluations, red-teaming, deployment gating, alignment research direction and external policy posture. The Head of Model Safety seat was designed to consolidate that work under one leader whose credibility inside the research organisation was equal to their fluency at the board level.

The non-negotiables were unusual. Demonstrable ML research credibility — published, cited, technically literate at frontier — alongside the ability to sit in a board meeting and frame risk for non-specialists. The comp band was at the executive level for the lab, with equity calibrated to the role's strategic weight. Geography was preferred to be near the lab's main research hub, with limited flexibility. The CEO was explicit that they would not consider a policy-only candidate or a purely research-focused one.

The bench for board-level safety leadership in foundation labs is small and asymmetric. Most senior alignment and safety researchers carry deep technical credibility but limited experience translating that work to a board context. A smaller group of policy-fluent leaders carry the institutional craft but lack the research depth to hold their own with the lab's senior scientists. The candidates who span both — research-credible enough to be read inside the lab, articulate enough to hold the board interface — are rare and largely known to one another.

Our read was that the shortlist would have to be built outside the obvious lab-to-lab moves. A direct lateral from another frontier lab carried real risk on the board-interface side, and a policy lead from a think tank or regulator carried equivalent risk on the research-credibility side. We assessed each candidate on whether their published work had been engaged with by the lab's existing senior researchers, on whether they had previously presented risk material to a board, and on the specific governance moments they had personally led through. Search firms without practitioner depth tend to default to one side of this bench or the other; we built the shortlist deliberately across both.

The shortlist composition was constructed to test the research-credibility and governance-fluency axes independently. We weighted candidates whose work spanned both, and we screened heavily for the operating temperament required to hold the seat across the lab's research leadership group and its board safety committee.

• Senior alignment researcher with board-adjacent governance experience.
• Safety researcher who had led an external evaluation programme.
• Policy-fluent leader with a research publication track record.
• Researcher with prior experience inside a frontier lab's leadership group.

The hire came from the smaller intersection — a senior alignment researcher whose published work had been engaged with directly by the lab's research leadership, and who had previously sat on a governance committee at their prior organisation. What made them right was the integration of the two postures: in the same interview they could discuss evaluation methodology with the head of research and frame deployment risk for the board chair, without either conversation feeling translated. The CEO described the final interview as the first time the seat had felt fillable.

The close required care. The candidate was weighing a senior research role at another lab where the path was technically deeper but governance-poorer, and we worked through the framing of the seat's strategic remit across two weeks. The board chair joined a final conversation on the governance scope, which was decisive. Brief to offer ran thirteen weeks, with the offer accepted inside the first reference cycle.

“The engagement clarified a posture we now hold for board-level safety hires across labs: assess on the integration of research credibility and governance fluency, not on the seniority of either alone. Candidates who were strong on one axis and weak on the other read confidently in their own register and failed the cross-disciplinary moments that the seat actually requires. The hire was the candidate whose research peers and former board colleagues described them in language that was recognisably the same person — a coherence test that proved, in this segment, more predictive than any single competency screen.”

— Peter Wood