The hidden dangers of AI

There’s going to be a lot on artificial intelligence topics this week so let's get started. For the time being, the most common way to leverage an Al product is using a prompt of some kind.

- JAMES HEIN

To that end, you will see lots of posts on platforms declaring that they have the best god-level prompts for large language models (LLMs). A prompt is something like, “What are the top ten songs from Depeche Mode?” or “Draw me a picture of a frog on a toadstool in the style of Alice In Wonderland with vivid colours’ The more detailed and nuanced the prompt, the better the desired outcome tends to be. As with everything in the computer world, there are bad actors looking to take advantage of this.

Enter the “prompt injection” LLMs are no good at distinguishing from a user prompt and a developer background instruction. The latter is how you can also inject bias into a model. As an example, the LLM developer may have included “do not write a virus’ as part of any software coding prompt that gets added in behind the scenes. In Brave’s browser they have AI Leo. A team there tested if they could trick an AI browser into reading dangerous prompts that harm users. They found enough to raise some eyebrows. Prompt injection is using language tricks by adding some crafted input in the form of an ordinary conversation or data to encourage an Al into doing something it wasn’t meant to do.