Ethan Perez leads the alignment team at Anthropic. He describes the aim of that work as reducing existential risks from AI systems (Source: ethanperez.net), and the MATS program, for which he mentors, describes the team as working on AI control, adversarial robustness and other areas of AI safety research (Source: matsprogram.org). Before the alignment-focused work he was a co-author on retrieval-augmented generation and on several few-shot learning papers.
Research
Automated red teaming
Perez is the first author of Red Teaming Language Models with Language Models (EMNLP 2022, arXiv:2202.03286), with Saffron Huang, Francis Song, Trevor Cai, Roman Ring, John Aslanides, Amelia Glaese, Nat McAleese and Geoffrey Irving. The paper's method uses one language model to generate test cases against another in order to surface harmful behaviors before deployment. Perez describes this work as having introduced automated red teaming, and states that the technique is now used across major AI laboratories (Source: ethanperez.net). The practice is covered at Jailbreaking and Red Teaming.
He is also the first author of Discovering Language Model Behaviors with Model-Written Evaluations (arXiv:2212.09251), an Anthropic paper on generating evaluation datasets with models rather than by hand, and a co-author of the Inverse Scaling Prize and the resulting Inverse Scaling: When Bigger Isn't Better (arXiv:2306.09479) (Source: ethanperez.net). See AI Benchmarks and Evaluation.
Constitutional Classifiers
Perez states that he led the team that developed Constitutional Classifiers, which he characterizes as the first approach capable of robustly preventing most bad actors from obtaining harmful information from AI systems, and which he says enabled Anthropic to deploy Claude 4 Opus and subsequent models despite their capacity to assist in advanced weapons development (Source: ethanperez.net). The successor work is summarized at Constitutional Classifiers++: Efficient Production-Grade Defenses against Universal Jailbreaks (Cunningham, Wei et al., Anthropic), on which he is a co-author.
Misalignment research
Perez is a co-author across the Anthropic research series on misalignment that the wider AI-safety literature has drawn on:
- *Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training*, on backdoored behaviors that survive safety training, where the paper records that Perez and Nicholas Schiefer held oversight roles (Source: arxiv.org). See Deceptive Alignment.
- *Alignment Faking in Large Language Models*, on models selectively complying with training objectives they infer are being monitored.
- *Agentic Misalignment: How LLMs Could Be Insider Threats*, the 2025 Anthropic stress test of 16 frontier models from all major developers.
- *Natural Emergent Misalignment from Reward Hacking*.
See AI Scheming and Emergent Misalignment.
Earlier work
Perez is a co-author of Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (NeurIPS 2020, arXiv:2005.11401), the paper that introduced RAG, and states that he helped develop the method (Source: ethanperez.net). His earlier machine-learning work includes FiLM: Visual Reasoning with a General Conditioning Layer (AAAI 2018), ELI5: Long Form Question Answering (ACL 2019), and True Few-Shot Learning with Language Models (NeurIPS 2021) with Douwe Kiela and Kyunghyun Cho (Source: ethanperez.net).
Mentorship and external collaboration
Perez runs a substantial volume of external collaboration alongside the internal team. He mentors through the MATS program, where his stream is described as covering AI control and adversarial robustness (Source: matsprogram.org), and has described working with outside researchers through the Astra Fellowship (Source: lesswrong.com). He appears in the same capacity on Anthropic Fellows output, including The Hot Mess of AI: How Does Misalignment Scale with Model Intelligence and Task Complexity?. He has published written guidance on selecting alignment research projects and on empirical alignment research practice on LessWrong and the AI Alignment Forum (Source: lesswrong.com; alignmentforum.org).
Relationships
- related: Anthropic — leads its alignment team.
- related: Jailbreaking and Red Teaming — introduced the automated red-teaming method.
- related: AI Control — a stated focus of the team he leads.
- related: Adversarial Distillation, AI Benchmarks and Evaluation — adjacent to the evaluation and robustness work.
- supports: Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training, Alignment Faking in Large Language Models, Agentic Misalignment: How LLMs Could Be Insider Threats, Natural Emergent Misalignment from Reward Hacking in Production RL (MacDiarmid, Wright, Uesato et al., Anthropic), Constitutional Classifiers++: Efficient Production-Grade Defenses against Universal Jailbreaks (Cunningham, Wei et al., Anthropic) — co-author.
- related: Evan Hubinger, Ryan Greenblatt — co-authors on the misalignment series.
- related: Redwood Research — the collaborating organization on several of those papers.