How Hackers Attack AI av Vetle Hjelle

Hackers can use various attacks like prompt injection to substantially alter AI behavior to their advantage. In this presentation we will cover some of the ways that malicious actors can achieve their goals through attacking a Large Language Model (LLM). There are already proven successful attacks against LLMs. We will cover attacks like direct and indirect prompt injection, what these attacks look like, and potential mitigation strategies. This presentation is fitting for technical personnel who integrate LLMs as part of their solutions, but can be adapted for a less technical audience that are curious about how they can use AI safely.