Mantas Mazeika

Mantas Mazeika is an AI safety researcher who builds benchmarks and methods for testing the safety, honesty and real-world capability of language models. His first papers, written with Dan Hendrycks in 2018, covered noisy labels and anomaly detection, and in 2023 he was one of the authors of DecodingTrust, an Outstanding Paper in the Datasets and Benchmarks track at NeurIPS. In 2024 he was first author of HarmBench, coauthored the WMDP benchmark, defended his thesis Toward Managing Catastrophic AI Risks at Illinois, and joined the Center for AI Safety (CAIS). He was first author of Utility Engineering and the Remote Labor Index in 2025, and in 2026 CAIS announced his appointment to the European Commission's AI Act Scientific Panel. He is a Research Scientist at CAIS.

Now at: European Union

Timeline

2018–2019

Noisy labels and outlier exposure

With Dan Hendrycks, Duncan Wilson and Kevin Gimpel, he coauthors a NeurIPS 2018 paper that trains deep networks on heavily corrupted labels by using a small trusted subset of the data. His next paper with Hendrycks and Thomas Dietterich, Deep Anomaly Detection with Outlier Exposure, appears at ICLR 2019.
2023

Catastrophic AI risks overview

An Overview of Catastrophic AI Risks, coauthored with Dan Hendrycks and Thomas Woodside in June 2023, sorts the dangers of advanced AI into four areas: malicious use, AI race dynamics, organizational risks, and rogue AIs. He also coauthors X-Risk Analysis for AI Research with Hendrycks in June 2023, an analysis of how to navigate tail risks.
2023

DecodingTrust wins at NeurIPS

He is one of 19 authors of DecodingTrust, a trustworthiness evaluation of GPT-4 and GPT-3.5 that covers toxicity, stereotype bias, adversarial robustness, privacy, machine ethics, and fairness. NeurIPS 2023 names it one of the two Outstanding Datasets and Benchmark Track Papers.
February 2024–March 2024

HarmBench and WMDP

HarmBench, a standardized evaluation framework for automated red teaming, appears in February 2024 with him as first author; it compares 18 red teaming methods against 33 target LLMs and defenses. In March 2024, he is also a coauthor of the WMDP benchmark, 3,668 multiple-choice questions that proxy hazardous knowledge in biosecurity, cybersecurity, and chemical security.
March 2024

A thesis on catastrophic risk

At the Siebel School of Computing and Data Science, he gives the final defense of his thesis, Toward Managing Catastrophic AI Risks, in March 2024. David Forsyth is committee co-chair and director of research, and the committee includes Svetlana Lazebnik, Bo Li and David Krueger of the University of Cambridge. He is an alumnus of Bo Li's Secure Learning Lab at the University of Illinois.
August 2024

Research scientist at CAIS

He joins the Center for AI Safety (CAIS) in 2024. In August 2024, he coauthors Tamper-Resistant Safeguards for Open-Weight LLMs, which introduces TAR, a method for building safeguards that adversaries cannot remove even after hundreds of steps of fine-tuning.
2025

AI values and remote work

In February 2025, he is first author of Utility Engineering, which finds that independently sampled preferences in current LLMs show high structural coherence and proposes methods to control these emergent value systems. The MASK benchmark, which he coauthors in March 2025, measures the honesty of language models separately from their accuracy. In October 2025, he is first author of the Remote Labor Index, a benchmark of real-world, economically valuable projects on which the best AI agent reaches an automation rate of 2.5%.
2026

EU AI Act Scientific Panel

In April 2026, he coauthors AI Wellbeing, a paper on measuring and improving the functional pleasure and pain of AIs. In June 2026, CAIS announces his appointment to the European Commission's AI Act Scientific Panel, an independent body of 60 members that advises the European AI Office and national authorities on general-purpose AI models. He is a Research Scientist at CAIS.