4 sources
Useful for the attack-and-evaluation side of alignment work, especially long-context jailbreak research and the measurement work that turns safety concerns into concrete tests.
Start HereMany-shot Jailbreaking
Researcher Profile
Editor reviewedYuntao Bai
Alignment via AI feedback (Constitutional AI)
Post-training and model-behavior researcher at Anthropic
Important for understanding how Anthropic’s assistant-training stack evolved from early RLHF into Constitutional AI and later robustness work around jailbreaks and behavior control.
Organizations
Anthropic
Labs
About This Page
This profile is meant to help you get oriented quickly: why this researcher matters, what to read first, and where to explore next.
Last reviewed
March 18, 2026
Known For
The ideas, systems, and research directions that make this person worth knowing.
01
Helpful and harmless assistant training
02
Constitutional AI
03
Behavioral robustness and jailbreak research
04
Alignment via AI feedback (Constitutional AI)
05
Constitutional AI: Harmlessness from AI Feedback
06
Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback
Start Here
Canonical papers, project pages, or repositories that anchor this profile.
Supporting Sources
Additional links that help verify and flesh out this profile.
Related Researchers
People worth exploring next because they share topics, labs, or source material with this profile.
5 sources
A high-signal person to follow for the part of alignment research that asks whether a model’s stated reasoning can actually be trusted and measured.
4 sources
Important because he sits near the boundary between alignment theory and concrete failure-mode discovery, especially jailbreaks, preference training, and behavior evaluations.
5 sources
A high-signal page for anyone tracking whether model reasoning traces are actually trustworthy, not just fluent explanations pasted on after the fact.
5 sources
A strong person to follow for how Anthropic moved from assistant training into more explicit evaluation work around model behavior, red-teaming, and chain-of-thought faithfulness.
5 sources
Worth following for the evaluation side of Anthropic’s alignment program, especially where model-written tests and public-input methods become practical tooling rather than just ideas.