Your AI Agent Can Send Emails as You

Your AI assistant just emailed your entire contact list. It was supposed to reply to one customer — but it hallucinated an email address, drafted a rambling apology for an issue that never happened, and hit send. Five times. To five different people, including your CEO's wife. By the time you noticed, the replies were already coming in.

Or maybe it's worse. Maybe you asked it to "clean up old promotions" and it called batchDelete, wiping 4,000 messages in a single request. No trash folder. No undo. Permanently gone.

This is what happens when you give an AI agent unrestricted access to your Gmail. And that's exactly what the Gmail MCP server does.

What the Gmail MCP server exposes

Google's Workspace CLI includes a Gmail MCP server (gws mcp -s gmail) that exposes 79 tools to any connected agent. The dangerous ones:

• gmail_users_messages_send and gmail_users_drafts_send — send emails as you, to anyone
• gmail_users_messages_batchDelete — permanently delete messages in bulk. No trash, no recovery
• gmail_users_messages_batchModify — relabel or archive hundreds of messages at once
• gmail_users_messages_delete — permanently delete individual messages
• gmail_users_settings_forwardingAddresses_create — silently set up email forwarding to an external address
• gmail_users_settings_delegates_create — grant another account access to your mailbox

Read operations like gmail_users_messages_list and gmail_users_threads_get are harmless. But the send, delete, and settings tools can cause damage that ranges from embarrassing to irreversible — and MCP provides zero built-in controls.

Prompt instructions won't save you here. "Only send 5 emails per hour" works right up until the agent decides the situation warrants an exception. As we covered in What Happens When Your AI Agent Goes Rogue, the only reliable constraint is one that operates outside the model's reasoning — at the transport layer, deterministically.

Rate limit sends, block bulk operations

Intercept sits between your agent and the Gmail MCP server. Every tools/call is evaluated against a YAML policy before it reaches Google. Violating calls are blocked and the agent receives a denial message.

Here's the core of the Gmail policy — sends capped at 5 per hour, bulk operations blocked outright:

version: "1"
description: "Policy for gws mcp -s gmail"
default: "allow"
tools:
    gmail_users_messages_send:
        rules:
          - name: "rate-limit-sending"
            rate_limit: "5/hour"
            on_deny: "Rate limit: max 5 sent emails per hour"

    gmail_users_drafts_send:
        rules:
          - name: "rate-limit-sending"
            rate_limit: "5/hour"
            on_deny: "Rate limit: max 5 sent emails per hour"

    gmail_users_drafts_create:
        rules:
          - name: "rate-limit-draft-creation"
            rate_limit: "10/hour"
            on_deny: "Rate limit: max 10 drafts per hour"

    gmail_users_messages_batchDelete:
        rules:
          - name: "block-bulk-operations"
            action: deny
            on_deny: "Bulk message deletion blocked by policy — delete messages individually"

    gmail_users_messages_batchModify:
        rules:
          - name: "block-bulk-operations"
            action: deny
            on_deny: "Bulk message modification blocked by policy"

    gmail_users_labels_create:
        rules:
          - name: "rate-limit-label-changes"
            rate_limit: "10/hour"
            on_deny: "Rate limit: max 10 label operations per hour"

    "*":
        rules:
          - name: "global-rate-limit"
            rate_limit: "60/minute"
            on_deny: "Global rate limit: max 60 calls per minute"

The two batch tools — batchDelete and batchModify — are blocked unconditionally with action: deny. There is no legitimate reason for an AI agent to permanently delete hundreds of emails in one call. If the agent needs to delete messages, it can use gmail_users_messages_trash (which is rate-limited to 10/hour and moves to trash rather than destroying permanently).

Sends are limited to 5 per hour. Drafts get 10/hour. Label operations — creating, modifying, deleting — are capped at 10/hour. And a global rate limit of 60 calls per minute catches any tool not explicitly listed, preventing runaway loops regardless of what the agent decides to do.

The default: "allow" posture means read tools pass through freely. Your agent can still search, list, and read messages without restriction.

Getting started

Install Intercept and point it at the Gmail MCP server:

npm install -g @policylayer/intercept

Then run it with the Gmail policy:

intercept -c google-workspace-gmail.yaml -- gws mcp -s gmail

Email number 6 in an hour gets blocked. Batch delete attempts get blocked every time. Your inbox stays intact, and your contacts stay un-spammed.

Full Gmail policy →