Year 2 Student - 2025 Cohort - Cohort 7

My academic background began in Cardiff Metropolitan University, graduating with a degree in Computer Security, my final year dissertation focusing on large language models, more specifically on exposing an attack vector in the training data through poisoning the dataset as well as how to detect these backdoors and how to mitigate them. During my studies, I achieved several certifications, including my ISO27001 and CompTIA Security+ which helped to strengthen my technical understanding as well as to build the governance and risk-management mindset behind security system design.

I chose to pursue a PhD because in my view emerging technologies such as machine learning have birthed security questions that our current understanding of the field simply cannot answer. These innovations are quickly becoming a part of the critical infrastructure we all rely on daily, but the security implications of how they are trained, deployed, and trusted remain underexplored. I wish to further my understanding of the field and shape how we defend our critical infrastructure before vulnerabilities become threats. The CDT offers a perfect balance between research depth, collaboration, and real-world impact, and I believe it’s a place where I can be a part of building the security standards that the future of technology will rely on.

Measuring and Improving the Fidelity of LLM-Orchestrated Synthetic Cybersecurity Telemetry

Defensive machine-learning systems that exist to protect modern infrastructure need to be able
to learn from security telemetry, yet the logs, network flows, and the endpoint events they depend on are exactly what organisations are unable to share for reasons of privacy, regulation, and infrastructure exposure. Public benchmark datasets are dated and demonstrably defective, and the synthetic alternatives that promise to be able to replace them are validated only through appeals to plausible appearance. The scientific gap is therefore one of measurement as much
as generation: no validated link exists between any measured notion of telemetry fidelity and the defensive utility that fidelity is supposed to deliver.

This project therefore aims to develop a multi-dimensional framework for measuring that fidelity and builds an LLM-orchestrated enterprise sandbox in which agents act as users, administrators, and attackers so that genuine systems emit telemetry natively, and determines experimentally which fidelity dimensions predict transfer to unseen data. The outcome closes
the loop between generation, measurement and utility, and releases a validated framework, reusable sandbox tooling and privacy-tested datasets.

June 14th 2026 - June 29th, Industry Placement with BT (Joseph Burton, Ayesha Abdul Majeed)

October 26th - November 9th, Academic Placement with Alma Oracevic

https://www.linkedin.com/in/dylan-jones-bab646387/