Playground

Try the guardrails yourself. Scan runs one guard stage over any text, with no model involved. Attack runs a poisoned page through the real agent twice, with no defence and with the guard on, against fake read-only tools.