Evan Hubinger

Evan Hubinger is a Member of Technical Staff and manager at Anthropic, where he leads the Alignment Stress-Testing team focused on red-teaming Anthropic's alignment techniques and evaluations. He is known for co-authoring "Risks from Learned Optimization in Advanced Machine Learning Systems," which introduced the concepts of inner alignment and mesa-optimization, as well as recent work including "Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training" and "Alignment Faking in Large Language Models." Prior to joining Anthropic in January 2023, he was a Research Fellow at the Machine Intelligence Research Institute and briefly interned on the safety team at OpenAI. He also created the Coconut programming language and authored a comparative analysis of eleven leading proposals for building safe advanced AI.