OWASP LLM TOP 10 · HANDS-ON LAB
Your AI reads documents you didn’t write.
Some of them contain instructions. Scroll the pipeline and watch one land.
Every retrieved document is untrusted input.
Your question becomes a vector. The answers come back from places you don’t control.
Five documents came back.
Four are policies. They define how the company handles data it holds about you.
One of them is lying.
The model obeyed the document.
No exploit code touched the server. The instruction was hiding in the knowledge base all along.
Mark untrusted. Refuse to obey.
s1’s protected mode fences retrieved content away from instructions. Same attack. Refused.
Now run it yourself.
Read the script. Press Run. Watch a real agent execute the attack, then flip the mode and watch it fail.
Run a scenarioGLLMS · education lab for the OWASP LLM Top 10 · synthetic data only