Dan Hendrycks is an American AI safety researcher who co-founded the Center for AI Safety (CAIS) in 2022 and serves as its Executive Director. He holds a PhD in Computer Science from UC Berkeley, has designed several widely used evaluation benchmarks, and also serves as a safety advisor to xAI.
Background
Hendrycks is a technical-safety researcher who has published on robustness, out-of-distribution (OOD) detection, machine ethics, and catastrophic risks. He co-founded CAIS in 2022 and co-designed evaluation benchmarks including MMLU (Massive Multitask Language Understanding), the MATH benchmark, and numerous robustness and OOD benchmarks. The "Hendrycks's law" attached to his name is the informal observation that algorithmic efficiency in frontier AI improves at a roughly constant compounding rate, often cited alongside compute and data scaling laws (see Scaling Laws).
Roles
Hendrycks is Executive Director and a co-founder of the Center for AI Safety. He also serves as a safety advisor to xAI. The xAI affiliation is described by some as a conflict of interest: CAIS positions Hendrycks as an independent safety voice, while critics point to Elon Musk's public posture on competing labs and safety discourse.
Positions and statements
Hendrycks argues for treating extreme AI risks — misuse, loss of control, and race dynamics — as first-order policy priorities comparable to nuclear and biological risk. He has emphasized benchmark-driven, evaluable safety over philosophical debate, and created the Weapons of Mass Destruction Proxy (WMDP) benchmark for dual-use capability evaluation. On regulation, he supported California SB 1047 and continues to advocate for transparency and testing requirements on frontier developers.
In May 2023 he organized and co-authored the CAIS Statement on AI Risk, signed by Geoffrey Hinton, Yoshua Bengio, Sam Altman, Dario Amodei, and hundreds of other senior AI figures.
Notable publications and works
- Introduction to AI Safety, Ethics, and Society (Hendrycks, 2024), a textbook used as a teaching resource in AI-safety curricula.
- MMLU benchmark paper (Hendrycks et al., 2020).
- MATH dataset paper (Hendrycks et al., 2021).
- "An Overview of Catastrophic AI Risks" (Hendrycks, Mazeika, Woodside, 2023).
- Organizer, CAIS Statement on AI Risk (2023).