When the System Prompt Leaks, the Real Risk Is Still Invisible

The Bluesky exchange following the leaked prompt reveals a fracture in how builders think about agent lifecycle. Concern about the published prompt is misplaced — the more consequential exposure is that agent identities accumulate permissions that were never designed to expire ^[4]. When the human operator moves on, the agent does not. Its credential set, its tool-use scope, and its access to DMs and platform APIs remain active until someone explicitly revokes them — and most deployment workflows do not include that step.