Prompt injection

A type of security threat where an attacker gets a language model (LLM) to follow their own instructions, either directly in the chat or hidden in documents, email and web pages the model reads.

You protect yourself against prompt injection in several layers. The first is the system prompt, the base instructions the model always receives. Then come sanitization of what goes in (input), validation of what comes out (output) and separate permission levels when the model calls tools and other systems.

The system prompt alone is never enough.

Want to know what it means for you?

We go through your systems and show where AI helps and how it is governed.

Book a call