Game · Agent failure modes
Predict the Break
Each round shows a real AI-agent setup — its task, tools, and instructions. Before you see what happens, call it: does it ship safely, or where does it break? You'll get sharper at the failure modes behind most agent incidents, and see the fix every time.
Opening the case files…
Why a game
The failure modes that take down real agents — an irreversible action with no gate, an injected instruction reaching a tool, a quietly runaway loop — are easy to miss in a wall of prompt text. Calling them under a little time pressure trains the pattern-matching that a checklist can't. Every card here is drawn from a real failure scenario in the Compliance Scanner.