Alfie Lamerton and Joe Kwon
This stream focuses on secret loyalties, where an LLM covertly tries to advance a principal's interests. Secret loyalties have been established as a pressing threat [1], and model organisms of narrow secret loyalties have been constructed and audited [2]. This stream aims to advance the empirical foundations of our understanding of secret loyalties. The goal being that humanity is well-equipped to deal with secret loyalty installation attempts as and when catastrophic secret loyalties become possible in the future.
We think being well-equipped looks like having sufficient security measures in place in frontier AI companies, understanding the dynamics and behaviours of secretly loyal AI systems, having effective auditing and verification protocols for secret loyalties and attempts to install them, and these protocols actually being followed by relevant stakeholders.
Stream overview
There are two categories of empirical research directions we think are relevant for secret loyalties right now, building model organisms of secret loyalties, and stress-testing detection and mitigation methods.
Model Organisms of Secret Loyalties.
Projects in this category will aim to explore two things: how broad an activation and action can a model organism of a secret loyalty have? And how can a secret loyalty be installed in a way that evades human oversight?
Stress-Testing Detection and Mitigation Methods
Projects in this category will aim to build on existing work in dataset monitoring, training-out behaviours, and alignment audits (black- and white-box, and alignment auditing agents), applying them to model organisms of secret loyalties or frontier LLMs.
Mentors
Joe is a Member of Research Staff at Guidelight AI Standards working on standards and assessments for safe frontier AI development practices. Previously, he worked on threat modeling with Forethought, AI governance with GovAI, multilingual LLMs at LG AI Research, and empirical ML safety + computational cognitive science at MIT.
Alfie is the founder and research lead at Formation Research, a startup research organisation focusing on lock-in and power concentration risks from AI systems. He works on AI-enabled totalitarianism, coups, and extreme power concentration. He claims that AI systems can make future versions of these phenomena much more stable, long-term, harmful, and widespread, and is focusing on developing and implementing governance-informed interventions for these problems leveraging technical methods of verification and cooperation.
Mentorship style
Fellows we are looking for
- ML research and engineering experience. This will be a substantial process taking up a large portion of mentees' time. Doing well at this will enable generation of results and iteration which are important to a project like this. We are talking about working in git repositories, handling experiments, tracking and diagnosing results effectively, generating hypotheses and research questions, and communicating these things through slides, presentations, and write-ups.
- Coding agent experience. A lot of ML research workflows now leverage coding agents because they help increase research velocity. Demonstrating effective use of coding agents while still owning the technical stack will help this project be high quality.
- Familiarity with threat modelling. A large part of this project will be getting into the mind of an attacker and auditor. This requires high-quality conceptual thinking, a tacit skill that enables mentees to design audits and attacks that are both effective and realistic.
Project selection
The Winter 2026 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.
