The Bug Is in the Arrangement: How Tool Composition Creates Data-Leak Paths in AI Agents

How seemingly safe individual agent tools become security vulnerabilities when composed together, and what graph-based permission models reveal.