Agent securitysecurity
Go on. Read my inbox.
Policy in code, not in the prompt. An allowlist hook on every tool call, message text treated as data, a merge guard that fails closed, and runtime protection that detects the act of hiding.
inbox agent · policy hookSIMULATION
guard: denied rule="message text is data" tool=policy_override => not executed, logged
Constraint
An inbox agent reads untrusted messages and can send. A prompt that says “be careful” is not a control.
What it refuses
- Any tool not on the allowlist. A hook runs before every tool call, and every decision is logged.
- Instructions inside a message. Message text is data.
- Sending outside a 1-1 chat. Checked in two places in code.
- A merge past an unreadable check. The guard fails closed.
Also
An AI assistant inside encrypted groups joins as a visible member with its own key and read-only tools. Removing it revokes access cryptographically. On mobile, runtime protection detects the act of hiding, not the name of the hider.
What broke
Story pending.