---
title: "Lock quarantined mail away from an agent"
description: "After a few rounds of feedback from users, a new permission, message_safe_read, gives an agent a key with no way to retrieve what a quarantined sender wrote. The agent reads the cleaned text of ordinary mail, and for quarantined mail only the sender's address."
canonical: "https://machinemail.dev/changelog/agent-safe-view"
date: "2026-11-16"
---

# Lock quarantined mail away from an agent

> After a few rounds of feedback from users, a new permission, message_safe_read, gives an agent a key with no way to retrieve what a quarantined sender wrote. The agent reads the cleaned text of ordinary mail, and for quarantined mail only the sender's address.

- Date: 2026-11-16
- Tags: API, Trust, MCP, Dashboard
- Source: https://machinemail.dev/changelog/agent-safe-view

The MCP tools already withheld quarantined mail from an agent by default. But the agent could still ask for it: `include_raw` returned the raw text and HTML, and over the API any key that could read messages got them too. We gave agents a lot of credit there, trusting them to leave that content alone unless a person asked for it.

After a few rounds of feedback with users, we learned that this is not enough for everyone. Some of you need quarantined text locked down completely, with no way for the agent to retrieve it. You can now give an agent a key that reads mail in the **agent-safe view** only. Choose &#x2A;*Agent (safe view)** when you create a key in the dashboard, or grant the `message_safe_read` permission through the API or the CLI. The OAuth scope is `messages.read.safe`.

- **Only the sender of quarantined mail.** A quarantined message shows the sender's address. Its subject, body, preview, attachment names and links are withheld on every route, and a person reviews it in the dashboard.
- **No way around it.** The key gets `403` for the raw message and the inbox export, cannot forward quarantined mail or remove its `QUARANTINE` label, and cannot create a key that reads more than it does. In the MCP tools, `get_message` refuses `include_raw`.
- **Cleaned text for everything else.** `text` is the trust layer's cleaned text, with quoted history, signatures and text hidden from human readers removed. Raw HTML, raw text and headers are left out.
- **Risky attachments are withheld.** Attachments of quarantined, spam, blocked or unauthenticated mail, and files rated high risk, are refused.
- **Events in the same view.** `message.*` events on the WebSocket, the event stream and in the event log reach the key in the agent-safe view. So do deliveries of a webhook the key creates or edits.
- **Existing keys are unchanged.** `message_read` is still the full view, and a key with both permissions gets the full view.

Anonymous sandbox agents now read mail in the agent-safe view until an organization claims them. After a claim they get every scope, as before.

A person who needs to see a quarantined message opens it in the dashboard, or uses a key with `message_read`.

Read more: [The agent-safe view](https://machinemail.dev/docs/concepts/api-keys#the-agent-safe-view) and [Create a key for an agent](https://machinemail.dev/docs/guides/api-key-for-an-agent).
