Securing LLM Use in Critical National Infrastructure-Adjacent Domains

Large language models are increasingly used in CNI-adjacent sectors such as law, journalism, mental health, and government. Misconceptions about privacy and data handling have already produced real-world harm, yet we have little understanding of how LLM interactions in these domains become exposed, misused, or exploited. This project will focus on three threat categories: accidental disclosure (logs, regenerated outputs, shared links), adversarial extraction (prompt reconstruction, conversation probing), and risky insider practices (copy-pasting sensitive information, poor mental models of privacy).

The student will map these risks and build a framework for understanding and mitigating them. The work will involve interviews, surveys, and co-design workshops with practitioners across these domains, supported by analysis of technical, legal, and UX safeguards. Expected outputs include a disclosure-risk framework, design guidelines, prototype interface concepts, and policy recommendations for organisations and regulators.

Required Skills

Applicants should have experience or strong interest in at least one of the following:

• human-computer interaction, privacy, or security
• AI governance, law, or policy
• qualitative research methods (interviews, surveys, workshops)
• foundational understanding of LLMs and their risks
• Technical skills are welcome but not essential.

Preferred Start Date

October 2026.