AI Applications Security Puzzle

Someone tells your AI assistant to ignore its instructions. It does. Someone asks it a question it was never supposed to answer. It answers. Someone crafts a message that makes it act outside its intended scope. And it doesn't push back.

These are patterns from live products, already deployed, already used. And in most of the teams where they happened, there was no test for it. Not because anyone was careless, but because nobody asked the right questions.

This talk is a grounded look at how AI systems get exploited in practice. We'll work through public cases, use the OWASP AI Testing Guide as our map, and name the failure patterns clearly enough that you'll recognize them when you see them in your own product.

As AI gets added to more products, the attack surface significantly grows. 44% is the year-over-year increase in the exploitation of public facing software. Most teams are focused on whether AI integration works: Does it give good answers? Does it feel right? Is it biased? The harder question: "What happens when someone tries to manipulate it?" — gets pushed to later. But later rarely comes.

AI security issues are not always caught by security specialists. They're caught by people who think critically, ask questions, and dig a level deeper to find the answer.

You may not walk out of this talk as an AI security expert. But you will walk out knowing what questions to ask and why asking them is already your job.