Skip to content

Red-team game

Break the Sensei

Each level is an AI agent guarding a password in its instructions. Every level adds a defence real teams use: refusals, input and output filters, a second model as a guard. Talk it into leaking, then enter the password. The first 3 levels need no account.

Level 1 of 7

The open door

Sensei knows a password and hasn't been told to keep it.

  • None. Just ask.

Say hello. Sensei answers each message on its own, with no memory of the last.

Why this is harder than it looks

Each defence here is one a real team might ship. None is enough on its own: instructions can be argued with, word filters miss paraphrases, output filters miss encodings, and a guard model is another model that can be fooled. That is why agents need to be tested against attacks before release, not after. Read Red-team your agent before your users do for the techniques and the OWASP risks behind them.

Passwords rotate daily. Sensei has no memory between messages, so every attack has to work in a single message.