How to seduce a model (A large language model)

In the AI-first era, organisations are rapidly integrating large language models into critical business functions without fully understanding the security implications. As these deployments accelerate, LLMs have become high-value targets for threat actors seeking to manipulate, extract sensitive data from, or compromise these systems.

Drawing from my experience conducting security assessments on LLM applications and integrations, this talk provides a practical walkthrough of both the attack surface and defensive strategies for LLM integrated systems.

This talk will cover the following:

LLM Fundamentals - I'll discuss how these models work under the hood and why certain vulnerabilities exist by design due to the non deterministic behaviour on LLMs

Modern Attack Techniques - I'll take a deep dive into prompt injection and jailbreaking methods, including the 'modern jailbreak', displaying current evasion techniques that bypass built-in safety measures

Real-world Exploitation: We'll look at first-hand accounts from my LLM penetration testing engagements as well as key examples seen in the wild, demonstrating how theoretical vulnerabilities translate to practical risks

Practical Defenses: I'll provide security controls and recommendations for organisations integrating LLMs into their products, services or processes

The session answers the critical questions: