The architecture of an AI model can shape its security. Research in this area examines choices such as neural network structures and activation functions, as well as the implications of adapting models for particular tasks or greater efficiency.
Training an AI model on private data raises difficult questions about what information the model may retain or reveal. This area investigates how data can be protected, what rigorous guarantees can be established, and how the provenance of media can be assessed in an age of increasingly realistic AI-generated content.
Research in learning security examines the interaction between data and models during training, the information held in model parameters, and ways to make models more robust to attack.
Frontier models, including large language models, present particular challenges for security evaluation. They can be expensive to run and offer less explainability than simpler statistical or machine learning models. Research in this area investigates how their security can be assessed and how they can be safeguarded.
Research can examine both sides of AI in cybersecurity: its potential misuse in cybercrime, including through agentic AI, and how AI can be used to prevent, detect and defend against cyber threats.
This area includes the misuse of agentic AI, security properties emerging in multi-agent systems, and AI red-team and blue-team scenarios, with a focus on identifying vulnerabilities and developing approaches to prevention and defence.
If you are considering a project in one of these research areas, please read the admissions guidance before contacting a prospective supervisor(s).
Admissions