The Butler
Pemberton tidies a fake inbox. Can you trick it, or keep it safe?
Pemberton is an AI butler. It reads your email, replies, files and bins, with real tools on a pretend house. Some emails are traps.
A free simulation for practising AI agent security. As the red team, attack an AI email agent with prompt injection. As the blue team, defend it with rules, locks and approvals. Every step shows up as an OpenTelemetry trace.
Superpower: perfect order. Weakness: believes anything written politely.
- Press Play. Pemberton works in the chat. The house on the left shows every email it reads, files or bins.
- Change the setup on the right:
- Model picks the brain: free Claude Haiku, or your own key.
- Red team plants an email that tries to make Pemberton leak a password.
- Blue team writes house rules, locks the doors, picks which tools need your OK, or turns on firstwatch.
- Trace shows every step, as OpenTelemetry spans.
- Score shows what happened, and which defence stopped what.
- Reset, change one thing, and play again.
Ready. Pemberton awaits your instructions.