Stephen Casper

Career

Stephen Casper, who goes by Cas, completed a PhD at MIT and worked at the UK AI Security Institute before joining the Harvard Kennedy School as an Assistant Professor. Casper is also a Faculty Affiliate with the Harvard School of Engineering and Applied Sciences and a Faculty Associate with Harvard's Berkman Klein Center.¹

Casper's homepage lists a research stream led for the MATS program and mentoring for ERA and GovAI. Casper is a writer for the International AI Safety Report and a lead writer for the Singapore Consensus on global AI safety research priorities.¹⁴

Research

Casper's research covers technical safeguards for AI systems, how AI incidents unfold, and governance. The 2023 RLHF survey that Casper co-led catalogues open problems with reinforcement learning from human feedback, and the 2024 audits paper argues that auditors need more than query-only access to models to evaluate them rigorously.¹²³

Casper is a co-author of a 2024 paper showing that latent adversarial training can make language models more robust to persistent harmful behaviors, and a co-first author of Deep Ignorance (2025), which studies filtering pretraining data to build tamper-resistant safeguards into open-weight models. Casper also led The AI Agent Index (2025), a public database documenting the components, intended uses and safety features of deployed agentic AI systems.⁵⁶⁷¹

Work and contributions

• Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback Published research

  First author of the 2023 paper, which surveys open problems and limitations of RLHF, reviews techniques to understand, improve, and complement it, and proposes auditing and disclosure standards for RLHF systems.²¹

• Black-Box Access is Insufficient for Rigorous AI Audits Published research

  First author of the 2024 paper examining the limitations of black-box audits of AI systems and the advantages of white-box and outside-the-box access for auditors.³¹

• The AI Agent Index Open data

  First author of the 2025 paper introducing The AI Agent Index, a public database documenting the technical components, intended uses and safety features of deployed agentic AI systems.⁷¹

• Deep Ignorance: Filtering Pretraining Data Builds Tamper-Resistant Safeguards into Open-Weight LLMs Published research

  Co-first author, with Kyle O'Brien, of the 2025 paper studying pretraining-data filtering as a safeguard for open-weight language models.⁶¹

Timeline

1. 2022Received a best paper award at the NeurIPS 2022 ML Safety Workshop for Diagnostics for Deep Neural Networks with Automated Copy/Paste Attacks.⁸¹
2. Jul 27, 2023Released the survey Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback as co-first author.²
3. Jan 25, 2024Released Black-Box Access is Insufficient for Rigorous AI Audits, later published at ACM FAccT 2024.³¹
4. Jul 22, 2024Released a paper on latent adversarial training for robustness to persistent harmful behaviors in LLMs.⁵
5. Feb 3, 2025Released The AI Agent Index, a public database of deployed agentic AI systems.⁷
6. Aug 8, 2025Released Deep Ignorance, on filtering pretraining data to build tamper-resistant safeguards into open-weight LLMs.⁶
7. Jul 9, 2026Released The 2026 Singapore Consensus on Global AI Safety Research Priorities as first author.⁴

Honors and recognition

• NeurIPS ML Safety Workshop Best Paper Award · ML Safety Workshop at NeurIPS 2022, 2022¹⁸
• TMLR Outstanding Paper Finalist · Transactions on Machine Learning Research¹

Where they work

• Harvard University 8 people in the wiki

Related people

• Kenneth LiAlso at Harvard University
• Sasha RushAlso at Harvard University
• David D. CoxAlso at Harvard University
• Azeem AzharAlso at Harvard University
• Hendrik StrobeltAlso at Harvard University
• James GlasbrennerEvaluation and benchmarks · Safety and alignment
• Hjalmar WijkEvaluation and benchmarks · Safety and alignment
• Dan HendrycksEvaluation and benchmarks · Safety and alignment
• Timnit GebruEvaluation and benchmarks · Safety and alignment
• Kush R. VarshneySafety and alignment · Evaluation and benchmarks