The Capability Paradox: How Smarter Auditors Make Multi-Agent Systems Less Secure