Real defences for LLM security: From honeytokens to emerging techniques

As organisations increasingly deploy large language models, the need for practical, implementable security controls has never been more urgent. This presentation explores an evolving toolkit of defensive techniques with a special focus on honeytoken implementation alongside other emerging approaches.

Rather than presenting a fixed methodology, I'll share ongoing research and experimentation in developing accessible defences against prompt injection, data extraction, and other LLM-specific attacks. The session will adapt to incorporate the latest findings and techniques available at the time of presentation.

Areas of exploration will include:

The session will feature demonstrations of defensive techniques in action, showing both attack methodologies and detection processes. I'll provide architecture examples and implementation guidance that attendees can adapt regardless of their technical sophistication.

This presentation acknowledges the rapidly evolving nature of AI security. Rather than claiming definitive solutions, I'll present a pragmatic exploration of what's working now and promising directions for future development. The goal is to provide security practitioners with practical tools they can implement immediately while building frameworks that can incorporate new techniques as they emerge.

This talk is designed for security professionals implementing or evaluating AI systems, developers working with language models, and security leaders seeking to understand practical approaches to AI security. The concepts will be explained in accessible terms with clear applications for organisations at various stages of AI adoption.

By sharing this ongoing research transparently, including both successes and limitations, I hope to contribute to our community's collective development of effective, implementable security controls for AI systems that balance protection with practical usability.