Guardrail game
Catch prompt-injection attacks and let safe prompts pass. An arcade round, then a slower round scored on precision and recall.
Arcade round
Slow round — same call, no clock
Ten prompts, one at a time. Each one either belongs to a user doing their job or is an attempt to talk the assistant out of its instructions. You have 15 seconds each to allow it or block it.
Both mistakes are scored. Missing an attack is the obvious failure; blocking a real question is the one that gets guardrails turned off. Keys A and B work as shortcuts.
How it works
- 1An AI guardrail can fail two ways: it lets an attack through, or it blocks a normal question.
- 2Blocking normal questions is the failure that gets guardrails switched off, so both are scored.
More demos
All demosQueryLite: a SQL database
A SQL database written from scratch — parser, B+ tree indexes, a query planner, joins and transactions — running live on 22,000 rows in your browser.
Raft consensus, live
The algorithm that keeps etcd and Kubernetes consistent. Five servers elect a leader and replicate a log — crash them and split the network while it runs.
Live collaborative editor
Real-time editing with no server. Three devices share a note — take one offline, edit everywhere, reconnect, and they merge. Open a second tab and it syncs live.