How to Securely Deploy Agents that Make Sensitive Decisions in Untrusted Environments