Sandipan Kundu

Sandipan Kundu is an Anthropic researcher whose path ran from theoretical physics to AI safety. He wrote on strongly coupled gauge theories, causality and chaos at the University of Texas at Austin, Cornell University and Johns Hopkins University. In December 2022, he was among the authors of Constitutional AI: Harmlessness from AI Feedback and of Discovering Language Model Behaviors with Model-Written Evaluations, and in October 2023 he was first author of Specific versus General Principles for Constitutional AI. Many-shot Jailbreaking, which he co-authored, appeared at NeurIPS 2024. He is a member of technical staff at Anthropic, focused on AI safety and the behavioral evaluation of language models.

Now at: Anthropic
Works: 38
Citations: 1,727
h-index: 19
Badges: Bridge

Publication numbers from OpenAlex, as of 22 October 2026.

Timeline

2012

Strongly coupled gauge theories

In the Theory Group of the University of Texas at Austin's Department of Physics, he writes with Willy Fischler in 2012 on the high and low temperature behavior of non-local observables in strongly coupled gauge theories that are dual to AdS.
2016

Causality and null energy

At Cornell University's Department of Physics, he writes Averaged Null Energy Condition from Causality with Thomas Hartman and Amirhossein Tajdini, posted in October 2016. The paper shows that microcausality makes the averaged null energy positive in interacting quantum field theories in more than two dimensions.
2018

Postdoc at Johns Hopkins

As a postdoc in the Department of Physics and Astronomy at Johns Hopkins University, he writes A Conformal Collider for Holographic CFTs in 2018 with Nima Afkhami-Jeddi and Amirhossein Tajdini. The paper builds a formalism for how causality constrains OPE coefficients in conformal field theories with large central charge.
2021

Saturating chaos bounds

Writing alone from Johns Hopkins University in 2021, he publishes Extremal Chaos. It finds a unique analytic extension of the maximally chaotic out-of-time-order correlator that saturates all the subleading chaos bounds which allow saturation.
December 2022

Constitutional AI and model-written evaluations

In December 2022, he is among the authors of Constitutional AI: Harmlessness from AI Feedback, which trains a harmless AI assistant through self-improvement with a list of rules or principles as the only human oversight. He is also among the authors of Discovering Language Model Behaviors with Model-Written Evaluations, in which language models generate 154 datasets and reveal new cases of inverse scaling.
2023

Faithfulness and general principles

In July 2023, he is among the authors of Measuring Faithfulness in Chain-of-Thought Reasoning, which finds that models produce less faithful reasoning as they become larger and more capable. In October 2023, he is first author of Specific versus General Principles for Constitutional AI, which asks whether models can learn general ethical behaviors from a single written principle.
2024

Many-shot jailbreaking

In 2024, he is among the authors of Many-shot Jailbreaking, presented at NeurIPS 2024. The paper studies long-context attacks that prompt a language model with hundreds of demonstrations of undesirable behavior, and finds that the attack's effectiveness follows a power law up to hundreds of shots.
As of October 2026

Member of technical staff at Anthropic

He is a member of technical staff at Anthropic, focused on AI safety and the behavioral evaluation of language models, after earlier research in theoretical physics.