Adam Gleave

Career

Gleave completed a PhD in artificial intelligence at UC Berkeley from 2017 to 2022, advised by Stuart Russell, with a thesis titled "Towards Trustworthy Machine Learning". During the PhD, Gleave was a research intern at DeepMind in 2019 and 2021, working with Geoffrey Irving and Jan Leike.³²

Gleave co-founded FAR.AI in January 2022 and serves as its CEO. Gleave also serves on the boards of METR (since May 2024), the London Initiative for Safe AI (since July 2024) and the Safe AI Forum (since December 2024).³¹²

Research

Gleave's research concerns making advanced AI systems verifiably safe, ranging from value learning and robustness in deep reinforcement learning to detecting hidden or deceptive behavior in frontier models. The 2019 paper "Adversarial Policies: Attacking Deep Reinforcement Learning" and the 2022 paper "Adversarial Policies Beat Superhuman Go AIs" studied how trained agents can be exploited by adversarial opponents.¹⁵⁶

Gleave's reward-learning work includes "Quantifying Differences in Reward Functions" (ICLR 2021) and the imitation library of clean imitation learning implementations. Later work with FAR.AI colleagues covers scaling trends in language model robustness (ICML 2025) and deception probes in reinforcement learning from verifiable rewards (ICML 2026).³⁷

Work and contributions

• Adversarial Policies: Attacking Deep Reinforcement Learning Published research

  First author of the 2019 paper, presented at ICLR 2020, which showed that reinforcement learning agents can be defeated by adversarial opponent policies.⁵³

• Adversarial Policies Beat Superhuman Go AIs Published research

  Co-author of the 2022 paper on adversarial attacks against Go-playing AI systems, an ICML 2023 oral paper that received a best paper award at the NeurIPS ML Safety Workshop 2022.⁶³

• imitation: Clean Imitation Learning Implementations Open-source code

  First author of the 2022 report describing the imitation library of imitation and reward learning algorithm implementations.⁷²

Timeline

1. 2017Began a PhD in artificial intelligence at UC Berkeley, advised by Stuart Russell.³
2. May 2019Posted "Adversarial Policies, Attacking Deep Reinforcement Learning" as first author.⁵
3. 2022Completed the PhD at UC Berkeley with a thesis titled "Towards Trustworthy Machine Learning".³
4. Jan 2022Co-founded FAR.AI and became its CEO.³
5. Nov 2022Co-authored "Adversarial Policies Beat Superhuman Go AIs" and released the imitation library report.⁶⁷
6. May 2024Joined the board of METR as a director.³
7. Dec 2024Became a director of the Safe AI Forum.³

Other documented roles

Roles the sources document that are not confirmed as current.

• Research intern (2019 and 2021), DeepMind³

Honors and recognition

• Best Paper Award, NeurIPS ML Safety Workshop (for "Adversarial Policies Beat Superhuman Go AIs") · NeurIPS ML Safety Workshop, 2022³

Where they work

• FAR.AI 3 people in the wiki
• METR 5 people in the wiki
• London Initiative for Safe AI 1 person in the wiki
• Safe AI Forum 1 person in the wiki

Related people

• Ian R. McKenzieAlso at FAR.AI
• Karl BerzinsAlso at FAR.AI
• Shengjia ZhaoReinforcement learning · Safety and alignment
• Paul ChristianoSafety and alignment · Reinforcement learning
• Wojciech ZarembaSafety and alignment · Reinforcement learning
• Jan LeikeSafety and alignment · Reinforcement learning
• Lilian WengReinforcement learning
• Beth BarnesSafety and alignment
• Edward BeechingReinforcement learning
• Mark RussinovichSafety and alignment