Abstract

Large Language Models are rapidly becoming part of modern applications, powering chatbots, copilots, AI assistants and enterprise workflows. But integrating an LLM into an application introduces an entirely new attack surface that traditional application security controls were never designed to address.

This session explores how attackers can exploit LLM-powered applications through prompt injection, jailbreaks, sensitive information disclosure, insecure output handling, excessive agency and weaknesses in retrieval-augmented generation (RAG).

Through practical attack scenarios and demonstrations, we will examine how seemingly harmless user inputs and untrusted external content can manipulate AI behaviour, expose sensitive information and influence downstream application actions.

The session will also explore defensive strategies, including input and output validation, trust boundaries, privilege separation, secure RAG implementation, AI guardrails, monitoring and adversarial testing.

Attendees will leave with a practical understanding of the OWASP Top 10 for LLM Applications and a repeatable approach to identifying and mitigating security weaknesses in AI-powered applications.

Key takeaways:
- Understand the evolving attack surface of LLM applications.
- Recognise and reproduce common prompt injection and jailbreak attacks.
- Identify risks involving RAG, sensitive information and insecure outputs.
- Implement practical defensive controls and security testing strategies.
- Build security into the LLM application development lifecycle.