Hayden Williams

Software engineer and consultant in Cyber Security and AI, with a specialism in adversarial attacks against LLM-based agentic systems, prompt injection, and AI security research.

My dissertation for my BSc (Hons) Computer Science for Cyber Security at Oxford Brookes University (2023–2026), Project Earworm, is an audio prompt injection evaluation framework targeting voice-enabled LLM agents. Built on a custom ReAct architecture using Python, FastAPI, LangChain, Docker, and PostgreSQL, it is one of the first systematic empirical evaluations of this attack surface, and was shortlisted for exhibition at the university's TechShow 2026. More details are available on the Project Earworm website.

My engineering background spans Python, Java, C, C++, and C# (.NET MAUI), with a strong foundation in Unix/MINIX systems programming, POSIX IPC, and CISCO networking (CCNA).

I have built everything from secure Unix daemons and multi-agent debate simulators to decision tree classifiers and cross-platform applications. I am a NCSC CyberFirst Advanced alumnus and have previously published on the trajectory of large language models for the Oxford Union.

I work as Associate in AI at Oxford Capital, engineering AI agents and Azure data pipelines and helping lead secure generative AI and automation strategy in a regulated venture capital environment. I also serve as Director of Strategy at the Oxford Union, where I directed the Union's flagship Artificial General Intelligence debate. Outside of work, I sing as a Tenor at Lincoln College (University of Oxford) Chapel Choir.

Through my consultancy, I work with organisations on four areas: AI security advisory for LLM and agentic systems, AI agent design and automation, practical AI training for technical and non-technical teams, and ongoing advisory retainers.