Prompt Injection

If you've explored the field of AI security, even at a surface level, before taking on this learning path or room, it's very likely you will have encountered the term prompt injection. That's for good reason: some of the earliest attacks on generative AI systems that attracted widespread attention from social media, industry, and academia were primarily prompt injection attacks. However, with its mass popularity, the term gets thrown around a lot, and, without any of the core fundamentals being established, it can be confused with other, very similar terms. This leads to confusion all around. This module aims to be the antidote to that confusion, with this room laying the groundwork. Begin your journey into understanding prompt injection right here!

Learning Objectives

• Understand how LLMs interpret context and why that makes them vulnerable to prompt injection
• Understand the fundamentals of what a prompt injection attack is
• Identify both direct and indirect prompt injection techniques and their real-world consequences
• Recognise how prompt injection can subvert trusted AI systems through untrusted inputs, such as documents or tools
• Apply learned techniques in a simulated environment to exploit a vulnerable LLM integration

Learning Prerequisites

This room (and the Prompt Security module) is part of the AI Security path. At a minimum, you should have all the required knowledge contained within the AI/ML Security Threats room.