The startup’s stealth exit points to a new security frontier: not just screening what AI says, but trying to catch risky agent behavior while it is still unfolding.