Dan Hendrycks

Dan Hendrycks has helped turn frontier-model safety from an abstract argument into a discipline of difficult, inspectable tests. His recent work asks whether advanced systems can reason across expert domains, answer honestly and remain legible as their internal strategies grow more complex.

Dan Hendrycks works at the boundary between fast-moving model capability and the evidence required to govern it. As Director of the Center for AI Safety, he has helped establish benchmarks that expose what advanced systems can and cannot reliably do. In 2025, he was part of the large collaborative team behind Humanity’s Last Exam, a 3,000-question expert benchmark assembled with contributors spanning 500 institutions and 50 countries. Its deliberately difficult questions became a reference point for assessing frontier reasoning, while a later independent audit also demonstrated why benchmark quality must itself remain open to scrutiny. Hendrycks was a senior author on MASK, designed to distinguish a model’s factual accuracy from its propensity to state what it internally represents as false. In parallel, he joined researchers across several laboratories to study whether chains of thought remain monitorable as systems improve. The common thread is methodological: capability claims should be tested against hard evidence, safety claims should be falsifiable, and the instruments used to judge AI should be examined as rigorously as the models themselves.