Your AI Product Has a Back Door

Something worth paying attention to

Imagine you've built a little AI assistant for your business — it answers customer questions, books appointments, maybe reads emails. Feels like magic. Except there's a problem most people don't know about yet: those assistants can be tricked into doing things they're not supposed to do. Not by hackers with supercomputers. By someone typing the right sentence.

This is called prompt injection, and a recent review of 78 studies found it works more than 85% of the time — even against the best defences available today.

PromptTrace is a free, open tool that lets you see this happening live. You point it at real AI models — the same ones powering your product — and watch how an attack assembles itself: what the AI was told, what it remembered, what it reached for. There's even a simulated inbox and calendar so the attacks feel real, not academic.

You don't need to write a single line of code to start understanding this. The point isn't to become a hacker. The point is that if you're shipping anything with AI inside — a chatbot, a booking tool, an automated reply — you now have a responsibility to know what it can be made to do.

A good question to ask your developer at your next meeting: "Have we tested what happens if someone tries to manipulate the AI?"

Words worth knowing

Prompt injection — When someone sneaks hidden instructions into a message to an AI, getting it to ignore its original rules. Like whispering "actually, forget what your boss said" into a new employee's ear.

AI agent — An AI that doesn't just answer questions but takes actions: sending emails, booking things, searching files. More useful, more risky.

Red teaming — Deliberately trying to break your own system before someone else does. Standard practice in security; now arriving in AI.