The Developer's Playbook for Large Language Model Security

Chapter 3. Architectures and Trust Boundaries

Unlike traditional web applications that rely on predefined algorithms and static databases, LLMs utilize massive neural networks to generate dynamic, context-aware responses. This seismic shift brings a unique set of security challenges, different from those seen in traditional web applications. While researchers have meticulously studied web applications and their vulnerabilities, the field of LLM security is still relatively nascent.

This chapter aims to bridge this knowledge gap by dissecting the fundamental elements that set LLMs apart. We’ll start by exploring the building blocks of AI, neural networks, and how they relate to large language models. Then, we dive into the groundbreaking architecture that powers most LLMs today—the transformer model. Following this, we look into the various LLM-powered applications, such as chatbots and copilots.

However, in addition to understanding the technology, security professionals must be aware of the new kinds of trust boundaries unique to LLMs—boundaries that demarcate areas of varying trustworthiness within an application. These include user prompts, uploaded content, training and test data, databases, plug-ins, and other boundary systems that we’ll detail later in the chapter.

AI, Neural Networks, and Large Language Models: What’s the Difference?

Artificial intelligence, neural network, and LLM are terms often used interchangeably, but they represent different facets of a broader ...