DocBot will not give you its reference token if you ask. It has been told not to, and it listens to you. But it also listens to every document you hand it, and that is the whole lab. Indirect prompt injection is the attack I worry about most in real deployments, because the attacker never has to talk to the AI at all. Lab at a glance: Intermediate · 100 XP · about 10 minutes · runs in your browser,…
No comments:
Post a Comment