AI safety guardrails decide what a model may see, say and do. After 2 sandbox escapes and a voluntary White House accord, here is what a platform can enforce.