Co-first author of Liars' Bench and co-scientific lead of Aletheia's Quest, an LLM lie detection competition with NDIF supported by Schmidt Sciences. For about three years he has split his time between leading Cadenza Labs and doing research. He is a MATS 3.0 & 3.1 alumnus, and his work has been supported by Open Philanthropy, LTFF and Manifund.

Since 2023, Walter has mentored for SPAR and Algoverse, contributed to UK AISI capability benchmarks via the AI Safety Engineering Taskforce, and is the main contributor to EleutherAI's elk library. Other publications include Cluster-Norm for Unsupervised Probing of Knowledge (EMNLP) and AI–AI Bias (PNAS), which was covered by several news outlets including Forbes. Before moving into AI safety and his PhD at KIT, Walter spent roughly seven years as a software engineer at several German companies.

Co-first author of Liars' Bench and co-scientific lead of Aletheia's Quest, a lie detection competition run by NDIF and Cadenza Labs with support from Schmidt Sciences. Speaker at FAR.AI's 2026 deception workshop, and mentor in SPAR and Algoverse.

Kieron holds an M.Sc. in AI from the University of Amsterdam (cum laude), with a thesis on LLM honesty representations, and is an ML4Good and Talos Fellow. Before pivoting his career towards AI safety research, he spent over five years founding and leading two startups, one of which is still growing.

First author of “Probing and Steering Evaluation Awareness in Language Models”, studying evaluation awareness with whitebox tools, as part of the Pivotal Fellowship. Co-author of "18 Applications of Deception Probes", which proposed using probes to augment debate and monitoring. Finetuned model organisms for the lie detection competition Aletheia's Quest.

Peter is investigating whether meta-models, such as Activation Oracles and Natural Language Autoencoders, collude with the models they monitor and hide scheming behaviour from human overseers. He has previously worked on language model introspection and automated mechanistic interpretability.

Prior to AI safety, he worked for seven years as a software developer and research engineer in weather and climate forecasting at the UK Met Office.

Sharan is a research scientist at Google DeepMind working on AGI safety. He worked with us for 2.5 years, first as a SPAR intern and later as a part-time contractor, and is now a core advisor: he attends our weekly meetings and co-supervises mentees in programs such as SPAR.

He is also a PhD candidate in AI safety at the University of Cambridge, a MATS 7.0 alumnus, and first author of Open Character Training, the first open implementation of character training (with Evan Hubinger and Nathan Lambert).