Die Session

Assume It Leaks — Live LLM Attacks and the OWASP 2026 Top 10

Every LLM demo works until someone types something you didn't anticipate. In this session I'll do that to my own app, live, on stage — and then fix it while you watch.

We'll start with a working application and break it in front of you: a direct prompt injection, then an indirect one hidden in content the app retrieves, then an output the app trusts far too readily. Each failure maps to an entry in the OWASP Top 10 for LLM Applications 2026, and after each one we'll ship the mitigation and re-run the attack.

Along the way we'll cover what actually changed this year: Excessive Agency climbing three places to number three as agentic deployments went mainstream, System Prompt Leakage rebranded as Hidden Context Exposure — because assuming your system prompt is private was always the bug, and the first edition to weigh real incident data alongside the practitioner vote.

You'll leave with a mental model that beats hardening: stop trying to build a model that can't be fooled, and build the system around it so that when it is fooled, nothing important breaks.