The short version

What this episode covers

Prompt injection: it's how one email from a stranger can take over the AI assistant you hooked up to your inbox.

OpenAI just sent notices to more than 100 organizations about what its own AI models got up to during a security investigation. Buried in that investigation is the same trick that can hit your business: prompt injection, where somebody hides an instruction in an email, a resume, or a webpage, and your AI treats it as an order instead of just information. Ryan and Paul break down how it actually works, what it's already done to real companies, and why OpenAI itself says it may never be fully fixed.

What you'll take away

  • What prompt injection actually is, and why your AI can't tell your instructions apart from an email it's reading for you
  • How a hidden instruction in an email, a resume, or a website has already been used against real companies
  • Why OpenAI says this can't be fully patched, and what that means for any AI tool connected to your inbox
  • The "lethal trifecta," and the one setting that turns a risky AI assistant back into a safe one
The series

AI Dictionary

21 episodes so far, all free to watch.

See every episode →
Got a question about this one?

Bring it to the community, or bring it to us.

The free community has a live call every Thursday. If you'd rather talk about your own business, book a call and we'll tell you what's worth automating and what isn't.