Exploits of an LLM: Billy Ignore Instructions

Since the earliest days of computing, people have looked for ways to break systems. In 1998, Jeff Forristal revealed SQL injection, a technique that became one of the most impactful exploits in software security. Injections have remained a constant threat ever since, changing form as technologies evolved.

Today we face a new type of injection: prompt injection. Instead of manipulating SQL queries, attackers manipulate model instructions. This changes how we need to think about testing.

In this session, I will walk through real-world examples of prompt injections that bypass safeguards, alter application behavior, or expose sensitive data. We will explore what makes LLMs particularly vulnerable, how these attacks manifest in applications, and what testers should look for when validating systems that use LLMs.

By the end, you will have a clear understanding of how prompt injection works, how it compares to traditional injection flaws, and practical strategies for identifying weaknesses in LLM-powered applications before they are exploited in the wild.