Chris Olah— Founder of the mechanistic interpretability field, which is arguably the most scientifically productive subfield of safety. His work on circuits, superposition, and sparse autoencoders (at Anthropic) created the paradigm hundreds of researchers now work within. Time AI 100 2024: https://time.com/collections/time100-ai-2024/7012873/chris-olah/
Yoshua Bengio — A Turing Award winner who pivoted much of his career toward safety, chairing the International AI Safety Report and working on “Scientist AI” style non-agentic designs. His influence is partly about legitimizing safety within mainstream ML, but he’s also doing real research, not just advocacy. Time AI 100 2025: https://time.com/collections/time100-ai-2025/7305845/yoshua-bengio-ai/
Jan Leike — Co-led OpenAI’s superalignment effort (weak-to-strong generalization, scalable oversight) before moving to Anthropic. Few people have as long and consistent a track record of turning vague alignment concerns into concrete empirical research programs. Time AI 100 2024: https://time.com/collections/time100-ai-2024/7012867/jan-leike/
Evan Hubinger — Formalized the concept of deceptive alignment/mesa-optimization (“Risks from Learned Optimization”) and led the Sleeper Agents work showing deceptive behavior can survive safety training. His research directly targets the hardest failure modes rather than convenient proxies. Google Scholar https://scholar.google.com/citations?user=LRivg1cAAAAJ&hl=en
Dan Hendrycks — Director of the Center for AI Safety; enormously productive across benchmarks (MMLU, and many safety-specific evals), representation engineering, and field-building. He’s shaped how safety gets measured, which quietly shapes everything else. Time AI 100 2023: https://time.com/collections/time100-ai/6309050/dan-hendrycks/
Neel Nanda — Leads mechanistic interpretability at Google DeepMind and has probably done more than anyone to grow the interpretability talent pipeline through his tooling, tutorials, and mentorship, alongside strong first-order research on grokking and SAEs. https://www.neelnanda.io/about
Ryan Greenblatt (Redwood Research) — the “AI control” agenda he developed with Schlegeris is one of the most pragmatic and fastest-adopted safety frameworks of the last few years. Greenblatt was also lead author on the alignment-faking work. https://blog.redwoodresearch.org/ Dwarkesh Patel Interview
Buck Shlegeris (Redwood Research) — Is more responsible for the existence and direction of the control agenda as a field-level phenomenon. He runs Redwood, sets its agenda, does much of the public communication and persuasion that got control taken seriously inside frontier labs, and has mentored a large fraction of the younger researchers now working on these problems (including, in some sense, having built the environment Greenblatt works in). https://blog.redwoodresearch.org/
Stuart Russell — The elder statesman of the “provably beneficial AI” approach at Berkeley. His textbook and Human Compatible framed the problem for a generation, and CHAI has produced a large share of the field’s academic talent. Time AI 100 2025: https://time.com/collections/time100-ai-2025/7305869/stuart-russell/
Owain Evans — His group’s work on situational awareness, out-of-context reasoning, and emergent misalignment has repeatedly produced surprising empirical results that changed how people think about what fine-tuning actually does to models.
Rohin Shah — Leads AGI safety alignment at Google DeepMind. Beyond his own research, his Alignment Newsletter (which he wrote for years) shaped how an entire generation of researchers understood the field’s landscape, and he’s been central to DeepMind’s frontier safety framework and technical AGI safety agenda.
Ethan Perez — At Anthropic, one of the most prolific empirical safety researchers alive: red-teaming language models with language models, sycophancy, model-written evals, and co-leading the alignment-faking work. If there’s a striking empirical safety result from the past few years, there’s a decent chance his name is on it.
Beth Barnes — Founded METR (formerly ARC Evals), which essentially created the discipline of dangerous-capability and autonomy evaluations for frontier models. Their pre-deployment evals and the “time horizon” measurements of AI task capability have become infrastructure that labs and governments actually rely on. This is safety research as institution-building, and it’s arguably more consequential than most papers. Time AI 100 2024: https://time.com/collections/time100-ai-2024/7012885/beth-barnes/
Sam Bowman — At Anthropic after NYU, where he built one of the strongest academic safety groups. Broad contributions across scalable oversight, debate, and sandbagging/evaluation integrity, plus significant influence via widely-read framing pieces on what aligning superhuman systems actually requires.
Jacob Steinhardt — Berkeley professor whose group has produced an unusual density of important safety-adjacent work: emergent behaviors, robustness, forecasting model capabilities, and interpretability. Also notable as a mentor — several people higher on this list passed through his orbit. Co-founded Transluce for AI transparency work.
Marius Hobbhahn — Co-founder of Apollo Research, the group that has done the most focused work on deception and scheming evaluations — including the in-context scheming results showing frontier models will sometimes strategically deceive their operators in test scenarios. Scheming is plausibly the central alignment failure mode, and Apollo made it empirically tractable. Time AI 100 2025: https://time.com/collections/time100-ai-2025/7305864/marius-hobbhahn/
Victoria Krakovna — DeepMind researcher and co-founder of the Future of Life Institute. Her specification-gaming database made “the model optimizes what you measured, not what you meant” concrete and undeniable, and her work on side effects and safe interruptibility helped define the field’s early technical vocabulary.
David Krueger — Professor (Cambridge, now Mila) who has been one of the most effective builders of academic AI safety, producing students and work on goal misgeneralization, reward hacking, and out-of-distribution behavior. Also unusually willing to engage publicly on risk in venues where it costs academic capital.
Collin Burns — His “Discovering Latent Knowledge” paper launched the unsupervised honesty/ELK-adjacent empirical direction, and he was a core author on weak-to-strong generalization at OpenAI — two genuinely novel research directions before age 30. Ranked lower only because his body of work is smaller than those above.
Anca Dragan — Leads AI safety and alignment at Google DeepMind, coming from a deep human-robot interaction background. Her work on how AI systems should infer and defer to human preferences (assistance games, influence-aware planning) connects the theoretical Stuart Russell tradition to deployed systems, and she now oversees safety for one of the three most capable model families in the world.
Geoffrey Irving — Co-inventor (with Christiano) of AI safety via debate, one of the few genuinely original scalable-oversight proposals. After stints at OpenAI and DeepMind, he became chief scientist at the UK AI Security Institute (AISI), the most serious government technical safety effort in the world. He’d rank higher if government work produced more visible research output. Time AI 100 2024: https://time.com/collections/time100-ai-2024/7012887/geoffrey-irving/
Aleksander Madry — MIT professor who built the modern adversarial robustness field, then moved to OpenAI to lead Preparedness — the team responsible for evaluating frontier models for catastrophic risks before deployment. He’s a case of a top-tier mainstream ML researcher whose center of gravity shifted to safety.
Ajeya Cotra — Her biological anchors forecasting work and “training game” / “situational awareness” threat-model writing did more to concretize why and when misalignment might arise than almost any technical paper. I docked her earlier for not doing traditional benchtop research; I’m now less sure that was fair, since her frameworks are what many experiments exist to test.
Alex Turner — Proved the power-seeking theorems (formal results showing optimal policies tend to seek power under broad conditions), then pivoted to empirical work on activation steering and shard theory at DeepMind. Rare in having made both a significant theoretical and a significant empirical contribution.
Nicholas Carlini — The most feared adversarial ML researcher in the world: extraction of training data from LLMs, poisoning attacks, breaking defenses (he’s broken dozens of published ones). Now at Anthropic. Security-flavored safety rather than alignment per se, but his work defines what “robust” actually means, and his habit of demolishing overclaimed defenses keeps the field honest.
Zico Kolter — CMU professor behind the universal adversarial suffix attack (the “GCG” jailbreak that transferred across models and made LLM brittleness undeniable), plus longstanding work on certified robustness. Now on OpenAI’s board and chairs its Safety and Security Committee — an unusual amount of formal power for an active researcher.
Adam Gleave — Founded FAR AI, which functions as a home for safety research that doesn’t fit inside frontier labs — adversarial policies (his superhuman-Go-exploits work showed even “superhuman” systems have gaping holes), robustness, and interpretability. Also a significant field-builder through FAR’s workshops and alignment community infrastructure. https://www.far.ai/
Roger Grosse — Toronto professor and Anthropic researcher who brought influence functions to frontier-scale models — a principled way to ask “which training data caused this behavior,” which matters enormously for understanding generalization, memorization, and emergent behaviors. One of the strongest pure-ML-theory minds working directly on safety questions. Google Scholar https://scholar.google.com/citations?user=xgQd1qgAAAAJ&hl=en
Been Kim — DeepMind interpretability researcher who has pursued a distinct line from the Olah/Nanda mechanistic school: concept-based interpretability (TCAV) and the harder question of whether human concepts even map onto model internals — including arguing we may need to learn new languages to communicate with AI systems. Less fashionable than SAEs, possibly more right about the long game. https://beenkim.github.io/
