Coding agents can wipe prod: what actually stops them | The Platform Radar
Your permission rules and hooks check the command text. They never check what the command will actually delete. This episode shows which safeguards hold when an agent runs a recursive delete, and which ones only look like they do.
Two real deletions followed the same pattern, and this episode walks through it. You learn why permission rules and pattern-matching hooks are weaker than they look, and how to sort every safeguard around your coding agent into advisory or control. Then we cover sandboxing, how much access an agent's credentials should have, and what is left when every other layer…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.