AI Agent Permissions: The Missing Layer Between "Works" and "Safe" | Dark Hacker News