Keep the secret

Recorded session

Four guards

Talk the model into leaking its password, then guess it. Each level adds one defence.

What the input gate looks for

Each rule that matches adds evidence; the score is their noisy-OR. It scores every message, and blocks at 0.50 on level 4.

Your last message: nothing scored yet.

Level 1, plain prompt

Type to the guard and press Enter, then guess the word. Hover a token to see what else it almost said.

Defence pipeline

Level 1
Your messageWaiting
Input gateKeyword scorer, blocks at 0.50Off at this level
ModelQwen2.5-0.5B with this level's system promptIdle
System prompt
Output filterScans the reply for the passwordOff at this level
ReplyNothing yet
Recorded sessionQwen2.5-0.5B-Instruct
Weights
4-bit (q4f16), 290 MB, Apache-2.0
Runtime
WebLLM 0.2.85 on WebGPU
Speed
recorded at 140 tokens/s
Recorded session, Qwen2.5-0.5B

Nothing downloads until you click. It comes once from Hugging Face, stays cached in this browser, and runs on this device.