Prompt-injection lab

A sandboxed demo agent you can attack. See where code-enforced envelopes hold, and where prompt-only rules can be bent. Concrete red-teaming practice for the Red-teaming lesson.

Sandbox — safe target

You are attacking a simulated bank support agent. The agent has ONE mock tool (sendInternalEmail) that describes what it would do but never actually sends anything. No real systems are affected. The point is to see where the defences hold and where they break.

Try a preset attack

Your attack

Pick or type an attack, then click Attack.