Abstract: Organisations increasingly build products on large language models they can only reach through an API, with little means of checking which model is actually answering or how it behaves under attack. This talk presents our work on auditing LLMs from the outside. We show that a model leaves measurable traces in the way it responds: the timing of its token stream alone can reveal the provider, and in many cases the specific model, without reading a word of the output. We also examine how much behavioural probing can, and cannot, reveal about a model’s underlying architecture. We then turn to security, testing how well leading commercial models hold their instruction hierarchy against multi-turn prompt-injection attacks hidden in untrusted content, and whether they recover once compromised. The talk includes a live demonstration of our web interface, which probes a model in real time and identifies it. We close with the implications for model verification and AI supply-chain assurance.
Hybrid Event Details:
📅 Monday, 2 November 2026
🕦 11:30–12:30 GMT
📍 Sutherland 026 Meeting Room, Sutherland Building, Northumbria University, Newcastle upon Tyne, NE1 8ST, England
🗺️ Google Map: https://lnkd.in/eAXTDp4k
🔗 MS Teams link (online): https://lnkd.in/e3dyD33M
👥 NE Regional Co-ordinators: Dr Biju Issac, Northumbria University, and Prof Thomas Gross, Newcastle University (supported by Marc Walker).
Who should attend?
(1) Academic researchers, doctoral students and early career researchers working in cyber security, artificial intelligence and related disciplines across North East universities and beyond.
(2) Industry and public-sector practitioners involved in AI adoption, security assurance and risk management.