AI agents need permission boundaries as tool use expands

An AI agent that only produces text is low-risk, but its risk profile changes the moment it calls a tool. Minor permissions — reading logs, summarizing tickets, searching docs — tend to creep into broader access without explicit boundaries.