Que fait secure-agent-design ?
Distilled from the security architecture Meta disclosed for Muse (Sept 2026): each agent runs in its own secure VM, a separate "sentinel" system checks every action before it leaves the VM, and the agent never sees real passwords or card numbers. Security framed as architecture, not promises, was the adoption unlock.
The patterns
Per-user isolated compute
Each user's agent runs in its own virtual machine: an isolated computer dedicated to that user. Compromise of one agent cannot reach another user's data. The VM keeps working in the background even when the user is away.
Sentinel: pre-action verification
A separate system checks every action before anything leaves the VM. The agent proposes; the sentinel verifies against policy (allowed apps, spending limits, sensitivity rules) before execution. No action with side effects bypasses this gate.
Credential isolation
The agent never sees actual passwords or card numbers. Purchases go through single-use virtual card numbers issued per transaction, so the real payment credentials never enter the agent's context and can't leak through prompts, logs, or model outputs.
Human approval gates
The agent asks before sensitive actions: sending an email, spending money, deleting data. Approvals should be specific (this action, this amount, this recipient), not blanket. Routine low-risk actions can run autonomously once the user opts in.
Least-privilege connectors
The user decides exactly what the agent can touch: which apps are connected, whether access is read or read-write, and can revoke anytime. Default to the narrowest scope that completes the job; expand only on explicit request.
Transparency
The user can see everything the agent has done: a complete, human-readable action log. Include "forget" controls so users can delete what the agent has learned about them.
Messaging it
Lead with the architecture ("runs in its own secure VM, every action checked before it leaves"), not with adjectives ("bank-grade," "military-grade"). Name the components (VM, sentinel, single-use cards) so technical users can verify the claims. Acknowledge the tradeoff honestly: broader access means higher stakes, which is exactly why the architecture exists.