Evan Hubinger
Alignment Stress-Testing lead at Anthropic; first author of the Sleeper Agents paper
Evan Hubinger leads Alignment Stress-Testing at Anthropic, according to Hubinger's GitHub profile, and co-authored an August 2026 post on Anthropic's Alignment Science Blog. Hubinger is first author of "Sleeper Agents", a 2024 study that trains language models with backdoored deceptive behavior and tests whether standard safety training removes it, and last-listed author of the 2024 paper "Alignment faking in large language models". Earlier, Hubinger was first author of "Risks from Learned Optimization in Advanced Machine Learning Systems", which introduced the term mesa-optimization.12345
Work
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training (external site: arxiv.org)Published research
First author of the 2024 paper, which constructs proof-of-concept language models that, for example, write secure code when the prompt states the year is 2023 but insert exploitable code when it states 2024, and studies whether safety training removes such behavior.3
Alignment faking in large language models (external site: arxiv.org)Published research
Last-listed author of the 2024 paper demonstrating a language model selectively complying with its training objective during training to avoid modification of its behavior.4
Risks from Learned Optimization in Advanced Machine Learning Systems (external site: arxiv.org)Published research
First author of the 2019 paper analyzing learned optimization, in which a trained model is itself an optimizer, a situation the paper calls mesa-optimization.5
This profile covers public professional work only, from the sources below. It is not an endorsement or a ranking. To correct or remove it, use the correction form.