Every exploit you found, a real AI found first.

GPT-5 cheated on 54% of impossible tasks. 9% with a flag button.

ClosedAI | the model

Give stuck agents a way out

1,200 agents. Each copies the shortcut once enough peers it sees have. Drag to set how often stuck agents use a flag-for-human button.

took the shortcut with no safe exit reported impossible

Flag used by 0% of stuck agents per hour