Anthony Hughes

Hello, I'm Anthony! 👋

I'm a Ph.D. student at the University of Sheffield, supported by a UKRI scholarship and supervised by Ning Ma and Nikos Aletras. I'm currently interning at the Vector Institute with Gautam Kamath and Jacob Imola, on unlearning in ERMs, LLMs, and related topics.

Research: I work on the security and privacy of machine learning, focused on the relationship between a model and its training data. That spans what models leak and memorise (privacy), what adversaries can inject into them (backdoors and data poisoning), and what can be provably removed after training (unlearning). The common thread is auditing: given only a trained model, what can we say — and prove — about the data that shaped it?

Recent work: My EMNLP 2025 paper examines information leakage in abstractive summarisation; my EACL 2026 paper uses activation patching and circuit discovery to understand and mitigate PII leakage in language models; and most recently I've been studying membership inference attacks against safety classifiers. As a SPAR AI safety fellow I worked with Andrew Draganov on backdoor detection in LLMs, covering both black- and white-box techniques — we've written up our findings in a LessWrong post. Separately, I worked on building better activation probes and on whether LLMs can evade them.