Francis Rhys Ward
Currently I'm focusing on projects which evaluate sabotage, sandbagging, and oversight subversion in frontier agents (aka, building control evals). Additionally, I'm interested in coming up with novel alignment or control plans.
Stream overview
Two concrete project proposals:
- Realistic benchmarks for evaluating oversight subversion
- A core bottleneck to scalable/automated oversight research is high-quality, realistic benchmarks for studying oversight subversion. To help build effective oversight protocols, this project aims to evaluate AI agents’ capabilities to autonomously subvert oversight mechanisms in realistic environments.
- Collusion in automated oversight regimes
- Past work typically assumes that the automated overseer is trustworthy, for example, because it is not capable enough to pose a threat. We aim to relax this assumption to study the problem that frontier AI systems might collude with one-another to subvert human control.
Mentors
Rhys is a researcher at LawZero. He recently finished his PhD, on formalising and evaluating AI deception.
Technically, his work involves both conceptual research, in the intersection of game theory, causality, and philosophy, in addition to empirical evaluations of frontier AI systems. Rhys is now focusing on control-style research and frontier LM agent alignment.
He is a member of Tom Everitt's Causal Incentives Working Group. Previously, Rhys has worked at the Centre for Assuring Autonomy, the Center on Long-Term Risk, the Centre for the Governance of AI, and the UK’s AI Safety Institute.
In this MATS cohort, he is looking to work on evals and control-style research.
Mentorship style
Fellows we are looking for
- Good communication -- it's important that you keep me updated with how you're doing and how I can help!
- Experience with LM experiments (e.g., fine-tuning).
- Clear thinker and writer.
Project selection
The Winter 2027 cohort offers a wide range of research streams led by experts across AI alignment, interpretability, governance, and safety. Each stream provides its own research agenda, methodology, and mentorship focus.
