LLM based AI has introduced huge shifts in the technology landscape in a very short amount of time, a consequence of which has been the immense pressure on organizations of all types to adopt and/or develop any and all things AI. This pressure has resulted in widespread usage of fledging technologies often with surprising capabilities that are not well understood in terms of their security implications or interactions with other data and APIs. You yourself might be under that very pressure now!

This talk will provide a case study of a real world LLM based app that is vulnerable to a variety of attack vectors that illustrate the challenges to account for when integrating today's LLM technologies into web application stacks as well as how to protect against them. We will walk through a full attack pathway that culminates in the combination of the LLM with SaaS API's in order to gain full control. 

Things then get weird as we explore the ways in which attack payloads can be obfuscated for delivery and how we will need to adjust some of our traditional security approaches in response.

No deep AI or LLM knowledge is required for the talk, an overview of LLMs and the general attacks against them will be provided. Basic knowledge of limericks is advised.