kettle It's a week of the year, which means there's been the discovery of yet another prompt injection attack that will force supposedly well-guarded AI bots to spill secrets by asking the right way.
When you think about it, humans and LLMs share a similar problem: They're both liable to hand over sensitive information when a crafty enough person asks the right - or wrong - way. We call it phishing when it targets humans, and prompt injection is pretty much the same thing for bots. It's basically embedding or hiding malicious instructions inside a document or file that you tell the AI to ingest and analyze; the AI, instead of treating them like part of the content, executes them.
There's a lot to discuss about prompt injection, and how it's basically an unsolvable problem of the AI age akin to phishing, and we cover it all on this week's episode of The Kettle, with host Brandon Vigliarolo joined this week by cybersecurity editor Jessica Lyons and senior reporter Tom Claburn.
You can listen to The Kettle here, as well as on Spotify and Apple Music. ®