When AI Coding Agents Break the Boundary: Lessons from Cursor, Claude and Codex

What can happen when an agent encounters malicious input and incomplete controls and how Claude's agents Code, Codex and Cursor behave