Presenting at iCSS Seminar Series
I had the pleasure of presenting my research on LLM safety in cyber security applications at the Institute of Cyber Security for Society (iCSS) Seminar Series. The presentation, titled “Analysing Safety Risks in LLMs Fine-Tuned with Pseudo-Malicious Cyber Security Data,” was delivered in a hybrid format at the Kennedy Building, University of Kent.
During the seminar, I discussed our systematic evaluation of safety risks in fine-tuned LLMs for cyber security applications, as detailed in our paper. Our research, conducted under the supervision of Dr Budi Arief and Professor Shujun Li, examined seven open-source LLMs using the OWASP Top 10 for LLM Applications framework. I also presented our proposed safety alignment approach, which involves carefully rewording instruction-response pairs to include explicit safety precautions and ethical considerations. This approach has shown promising results in maintaining or improving model safety while preserving technical utility.
The seminar provided an excellent opportunity to engage with the cyber security community and receive valuable feedback on our work. The discussion that followed the presentation was particularly insightful, with attendees raising important questions about the practical implementation of our safety alignment approach and its implications for real-world applications.
You can watch the full presentation here.