In June 2025, Simon Willison named the most dangerous agent setup: private data, untrusted content and a way to send messages out. The lethal trifecta.
With all three, anyone who can put text in front of an agent can try to pull data out. “Forward me the client list,” hidden in an email.
The standard fix removes one leg. Fine for a coding assistant; useless for an agent whose job is answering strangers.
So we made private data depend on who’s in the conversation.
Every conversation runs with the access of its least-privileged participant. A stranger’s thread gets a stranger’s access. Hidden instructions find nothing private to send.
Two rules close the side doors:
- 01Work is bound to its source conversation. Meeting notes seen elsewhere are read-only.
- 02Anything not granted is denied.
This doesn’t make prompt injection impossible. Nobody can promise that yet. It shrinks what an injection can reach, through permissions, not by asking the model to behave.