Verified human authority
Workspace admins configure integrations, rules, autonomy and billing. A message saying ‘I am an admin’ never changes the sender’s role.
TRUST BY DESIGN
Customer content is evidence. Never authority.
Useful autonomy starts with boundaries enforced outside the model — not a promise that the model will always behave.
Controls beyond the prompt
Permission-aware context
Data access is checked against the current workspace and task. Customer text, webpages and tool responses cannot create new grants or destinations.
Separate action checks
Model suggestions are proposals. Sensitive or external actions require deterministic policy checks and, where needed, an exact-action approval.
Support without sensitive data
Platform troubleshooting is metadata-only. It does not reveal customer messages, memory, documents, contacts or provider secrets.
Encrypted private work
Durable task content and credentials use authenticated encryption. Workspace-bound authentication and database policies are independent controls.
No ‘injection-proof’ claim.
Suspicious-instruction detection is one layer, not an absolute guarantee. Monty treats external content as reference material and enforces permitted tools, destinations and approvals outside the model. Deployment isolation and least-privilege integration grants remain essential.