LLM Output Handling and Privacy Risks

Introduction

Large Language Models (LLMs) have transformed how applications handle data. From customer support chatbots to automated code review tools, they process and generate huge amounts of information. However, with this convenience comes new risks, and two of the most common are improper output handling and sensitive information disclosure. These issues fall under the OWASP Top 10 for LLM Applications 2025 (opens in new tab) as LLM05: Improper Output Handling and LLM02: Sensitive Information Disclosure, and they are becoming increasingly critical to understand when testing or building systems that rely on LLMs.

Learning Objectives

This room focuses on the risks introduced after an LLM generates its response. By the end of the room, learners will be able to:

• Understand how improper output handling can be abused to perform downstream attacks.
• Identify common cases of sensitive data leakage from LLM responses.
• Recognise how output can be chained with other vulnerabilities to escalate attacks.
• Apply defensive strategies to mitigate these risks in real-world applications.

Prerequisites

Before starting, it's recommended that learners have a basic understanding of:

• Web security fundamentals, including input validation and injection attacks.
• LLM basics, particularly prompts, system instructions, and context.