News
- Oct 2026Started as Principal Investigator on evaluation awareness in LLMs, leading a 4-person team on a $1M Coefficient Giving grant.
- Aug 2026My MSc thesis, Fuzzing Large Language Models to Elicit Hidden Behaviours, accepted at the COLM 2026 Actionable Interpretability Workshop.
- Jul 2026Most Current Model Organisms Are Leaky accepted at COLM 2026 — equal-contribution first author with Luca Baroni and Dan Wilhelm. An investigator agent using perplexity differencing set a new state of the art on AuditBench (74% vs. 59%).
- Jun 2026Awarded $1M by Coefficient Giving to study evaluation awareness in LLMs.
- May 2026Finished my MSc in AI at ARU with a Distinction (75.6%) — top student across the Computing department's master's programmes.
- Mar 2026Model Organisms Are Leaky submitted to COLM 2026 — perplexity differencing recovers the hidden finetuning objective across 76 public model organisms.
- Feb 2026MATS 10 exploration phase with Neel Nanda — steering with "intention vectors" induced emergent misalignment before any finetuning (slides).
- Nov 2025
- Oct 2025Started as a Visiting Researcher at Meridian, Cambridge, as Principal Investigator on a £230K Coefficient Giving grant to red-team sleeper agents and build better model organisms of misalignment.
- Sep 2025Participated in ARENA 6.0 at LISA; capstone with Luca Baroni: a model organism of encoded reasoning.
- Aug 2025Posted Mechanistic Exploration of Backdoored LLM Attention Patterns on arXiv.
- Jul 2025Joined MARS 3.0 with Stefan Heimersheim — computation in superposition via fact retrieval in real LLMs.
- May 2025Wrote up an analysis of attention patterns in data-poisoned sleeper agents from my master's research, and started as a Research Assistant at ARU on great crested newt re-identification.
- Jan 2025Pivoted into AI research — began an MSc in AI at Anglia Ruskin University (ARU).
- Jan 2023Kurdivia raised a pre-seed round from LoyalVC and Flat6Labs.
- Summer 2021Co-founded Kurdivia, a live trivia app that reached 1.4M downloads and 150K concurrent players.