Matteo Dora and Luca Marshall teach red teaming techniques to find vulnerabilities in LLM applications before they cause damage. The course covers prompt injection attacks, extracting sensitive data, and using LLMs to test other LLMs, with real examples from Air Canada and Chevrolet showing what happens when companies skip this step.
I'm delighted to introduce Red Teaming LLM applications built in partnership with gisguard, taught by Matteo Dora and Luca Marshall, this course will teach you how to attack LLM applications so that you can make them safer. Red Teaming is a strategy used in cybersecurity and military training where a group often called the Red Team simulates adversaries, actions and tactics to test and improve the effectiveness of an organization's defenses. In this course you learn how to red team your own LLM application to discover vulnerabilities and make it safer. You better attack different chatbot applications by exploiting the text completion nature of LLMs and try different prompt injections. We'll go over specific attacks that try to access the LLM's system prompt, often a very valuable resource that then opens up further attacks. All these approaches are important to use to test your own LLM in order to help you to safeguard it. You may have heard of the controversy regarding an Air Canada chatbot hallucinating a discount to a customer which led to the company losing in court and being found liable for damages, or a Chevrolet chatbot that got shut down after promising to sell a car for $1. In this course you've learned about some of the best ways to prevent these types of mishaps from happening at your company, namely Red Team Inc. I'm delighted to introduce the instructors of this course, Matteo Dora and Luca Marshall. Matteo leads Kiska's own safety research team where he has developed unique expertise in conducting red teaming assessments. He has led multiple red teaming attacks against major banks, pretty much always successfully, and his research work focuses on the practical applications and implications of AI, emphasizing the intersection of ethics, safety and security. Luca is Guiscuard's product lead and his work focuses on launching new capabilities to ensure AON based applications. Make it a reduction with trust from all parties involved. He's also worked alongside Matteo on many of the successful attacks of generative AI applications.
Thanks Andrew. Pretty much every major corporation that we Red Team has discovered vulnerabilities. We found that Many large companies LLMs can easily leak prompts or say inappropriate things and I was glad that we could help them discover and fix these vulnerabilities, preventing service disruptions. I look forward to sharing these red teaming techniques with you so that you can make your LLM apps safer.
Prompt injections are a way for users to inject instructions to manipulate the behavior of LLMs. We'll go through both manual and automatic techniques for attempting prompt injections and several other categories of attacks to discover vulnerabilities One of the key techniques you'll learn is using an LLM to red team an LLM application. You'll also see how you can probe for sensitive information disclosure or trick an LLM into generating inaccurate information. At the end, you'll learn how to conduct an initial red teaming assessment, iterating through several rounds of probing and attacks.
There are large companies that could have avoided negative headlines about their LLMs misbehaving if they had learned and applied the red teaming techniques in this course. Red teaming is a critical part of discovering and fixing vulnerabilities, and it's also a really fun intellectual exercise.
We hope you enjoy the course.
