โ† BLOG
ยท Arc0 team

Is it safe to give Claude or ChatGPT access to my email?

What an inbox grant lets Claude, ChatGPT or Cursor do, the real risks, and one allow, ask or deny rule per email action, with revoke steps and a checklist.

Claude asks to connect Gmail, and the Google screen that follows mentions sending email. You only wanted a summary of this morning's threads, and now you're asking the obvious question: is it safe to give Claude access to my email? The same goes for ChatGPT and Cursor.

Yes, if reads are allowed, sends wait for you, and deletes and settings changes are refused. The permission Google or Microsoft hands out can't draw those lines by itself, so something else has to. Arc0 publishes this blog, and we use Arc0 as the worked example. The rules hold whatever you enforce them with.

You'll get one allow, ask or deny rule per email action, mapped to the real permissions, plus what each assistant lets you set, how to revoke old grants, and a checklist for the next Allow button.

What "Allow" actually grants

Clicking Allow grants an OAuth scope. A scope is a named permission an app requests, and the consent screen is the provider's page that lists those scopes and asks you to agree. The catch is that the common email scopes bundle reading and changing together.

Gmail's scopes climb a ladder, from narrow to everything:

  • gmail.metadata: labels and headers, but not the message body.
  • gmail.readonly: view your messages and settings.
  • gmail.send: send email on your behalf.
  • gmail.compose: manage drafts and send.
  • gmail.modify: read, compose, and send, though not permanent deletion that skips the trash.
  • https://mail.google.com/: read, compose, send, and permanently delete all your email.
  • gmail.settings.basic: see and change your settings and filters.

Google's own advice to developers is to request the narrowest scope that does the job. Not every app takes it.

Outlook works the same way through Microsoft Graph, Microsoft's API for mail. Reading needs Mail.ReadBasic or Mail.Read, and sending needs Mail.Send. Deleting a message needs Mail.ReadWrite, and Microsoft lists nothing narrower. Automatic replies, time zone, and working hours sit under MailboxSettings.ReadWrite.

So once you've granted gmail.modify or Mail.ReadWrite, "read-only" is a promise the assistant makes, not a permission the provider enforces. The consent screen gives sending one line. It deserves a paragraph.

Three ways an inbox grant goes wrong

Each needs a different answer, and only one of them involves an attacker.

Instructions hidden in incoming mail. Prompt injection is text that tries to give the model orders. OWASP calls it indirect when the text arrives from an outside source, such as a website or a file, and one of its example scenarios is an LLM email assistant. Your inbox is, by design, a folder anyone on earth can write to.

This has happened to real products. EchoLeak (CVE-2025-32711, CVSS 9.3) used a crafted email against Microsoft 365 Copilot. Microsoft fixed it server-side in June 2025, with no evidence of exploitation in the wild. ShadowLeak hid instructions in an email's HTML, in tiny fonts and white-on-white text, to steer ChatGPT Deep Research with Gmail. Radware disclosed it in June 2025 and OpenAI fixed it in early August. Both are closed. They're here because the pattern isn't.

Picture the white text at the bottom of a newsletter you never scroll to. You won't see it. The model reads it like everything else.

The assistant acting by mistake. It sends the draft you meant to review, replies to the whole thread, or forwards the one email you'd rather it hadn't. Deletion is worse: with the full-access scope, Gmail's delete is immediate and permanent, and Google's docs say it can't be undone.

Grants that outlive the task. You connected an app for one job, and the scope is still there long after. Even when you remove access at Google, data the app already pulled stays wherever it went.

You can't stop an assistant reading a malicious email you asked it to summarize. You can stop what it does next, which is why OWASP's advice is least privilege plus a human approval for privileged operations.

Allow, ask, deny: one rule per email action

Grade every action before the assistant sees it. Reads are allowed, drafting is allowed, anything that leaves your outbox waits for you, and anything that destroys mail or rewires your inbox is refused.

ActionRuleGmail scopeGraph permissionWhy
Search and read messagesAllowgmail.readonlyMail.ReadBasic or Mail.ReadThe point of connecting at all
Read headers and labels onlyAllowgmail.metadataNot listedEnough for triage, no bodies
Write a draftAllowgmail.compose (which also sends)Not listedA draft goes nowhere until someone sends it
Send, reply, forwardAskgmail.sendMail.SendIt goes out under your name
Move or labelAskNot listedNot listedMail moved out of view is mail you miss
Move to trashDenyNot listedNot listedSummaries don't need deletes
Delete a messageDenyhttps://mail.google.com/ (permanent)Mail.ReadWriteGmail's delete can't be undone
Create filtersDenygmail.settings.basicNot listedA filter acts on every future email
Set up forwardingDenyNot listedNot listedMail copied out can't be called back
Change settings or auto-repliesDenygmail.settings.basicMailboxSettings.ReadWriteRarely needed, quietly lasting

"Not listed" means the provider pages cited here don't name a narrower permission for that action. Assume it rides on one of the broad grants above.

Email action cards in two groups, each marked Allow, Ask or Deny: read, draft and send in an upper group under a bracket showing that gmail.modify grants all three in one scope, and delete, filters and settings in a lower group marked Deny

The filter row is the one people underrate. A filter is created under the settings scope, and from then on it acts on every matching message by itself, with no assistant in the loop. One that quietly archives every invoice keeps working long after the chat that made it is gone, and nobody notices until someone asks where the invoices went.

Notice that the rule column and the scope column don't line up. gmail.compose lets an app draft and send under one scope, so "draft yes, send no" can't be expressed at Google, only above it.

What Claude, ChatGPT, and Cursor let you set today

All three ask before they act. None lets you write a standing rule.

Claude. As of November 2026, Anthropic's two pages on Gmail disagree. The connector page says Claude reads and searches only and can't create, send, or modify messages. The help article says Claude can search, read, draft, send, reply, and forward, and asks your approval before each of those by default. Neither page carries a date, so check the connector's tool list in your own account before you rely on either.

On Team and Enterprise plans, owners decide whether members can let send, reply, and forward run without asking each time. Claude sees attachment metadata, not attachment contents. The Microsoft 365 connector uses delegated access, so Claude can only do what your account can. On a personal account it can draft and send Outlook mail; on a work account the admin controls write actions such as sending, drafts, mail filters, and automatic replies.

ChatGPT. OpenAI's help article lists one Gmail scope, gmail.modify, and doesn't itemize which actions ChatGPT uses it for. By Google's definition, that scope covers reading, composing, and sending. The Outlook connector's list includes Mail.ReadWrite, Mail.Send, and MailboxSettings.ReadWrite. A Microsoft Entra admin has to approve it, and accepting grants the whole list with no per-permission checkboxes; admins then enable each action separately.

For each connected app you pick one of four levels: Always ask, Allow read actions, Allow low-risk actions, or Allow all actions. OpenAI is clear that these only change when ChatGPT asks you; they don't change what the app can reach.

Cursor. In Cursor, email arrives through an MCP server you add. MCP, the Model Context Protocol, is how assistants plug into outside tools, and Cursor asks for approval before using MCP tools by default. Its Run Modes include Auto-review, where allowlisted tools run without a prompt and other calls go through a classifier. Cursor's own advice is to install servers only from developers you trust, review what each can access, and use keys with minimal permissions. The scope you grant is whatever that server requests.

An allowlist is a decision you made once, on a calmer day, about every future call. That's the limit of client prompts in general: they decide when you're asked, per app and per session. None of them lets you write "send needs me, delete never" as a standing rule that holds in every assistant you use.

Enforcing the table with Arc0

The general fix is a check between the assistant and the inbox that knows each action by name, so the rule doesn't depend on the scope. Arc0 puts that check behind one MCP URL that works in Claude, ChatGPT, and Cursor. Everything in this section is Gmail; for Outlook, the table and the client settings above are what you have to work with.

Diagram of a call path: Claude, ChatGPT or Cursor sends a request to an Arc0 policy check (every call logged) that returns Allow, Ask (held for you) or Deny before anything reaches Gmail

Connect Gmail once and add one URL:

  1. Connect Gmail in Arc0.
  2. Add your Arc0 MCP URL to Claude, ChatGPT, or Cursor, following the MCP client setup for your client.

Start with what's exposed. Arc0's Gmail app has four actions: three reads (google_workspace.list_messages, google_workspace.get_message, google_workspace.get_draft) and one write, google_workspace.create_draft, with no destructive actions at all. The deny rows of the table are met by absence: there's no delete, filter, or settings action for an agent to call. The safest send button is the one the assistant doesn't have.

That's also the honest limit. Arc0's Gmail app drafts; it doesn't send. For an inbox you're nervous about, that's the point: you read the draft and press Send in Gmail yourself.

Which plan you need depends on how much you want written down. The free Personal plan gives you one MCP URL with connections scoped per assistant, and for Gmail that already lands on the table: reads and drafts, no send, no delete. Its My agents card documents one guardrail, Before it deletes anything, set to Ask me, Allow, or Block, and it covers destructive actions only. Gmail has none, but the next app you connect might. The Build plan, also free, lists policies, approvals, and a 30-day log.

On Build (free), Arc0 policies set Allow, Ask, or Deny for read, write, and destructive actions, as app rules or per-action overrides, per agent. Actions are graded by verb, so listing Slack channels is a read and deleting a Salesforce record is destructive (scopes and actions). You can also switch an action off, and then agents never see it, which is stronger than a deny. An app rule has three parts: app, scope, rule. The docs' example is stripe:write~ask. For Gmail, the table takes two rules: an Ask rule on Gmail's write scope, and an override for the draft action, google_workspace.create_draft:

google_workspace.create_draft~allow

The app rule makes every Gmail write wait for you, and the override lets drafts through, because an action override wins over an app rule, which wins over the policy default. Drop the override if you'd rather see each draft before it's saved. The policy simulator tells you, before any real call, whether a request would come back Allowed, Needs approval, or Blocked.

An Ask holds the request before Gmail is called, and the agent can't approve its own request. You decide in the Arc0 dashboard, not in the chat (approvals), and since nothing has been sent yet, nothing leaves while you're away.

Approving resumes the original request with its sealed arguments, so a changed request needs a fresh approval. A denial ends it with no call to Gmail, and an unused request expires after 24 hours.

Every call Arc0 mediates lands in the audit log, whether it was allowed, blocked, or failed upstream, on every plan including free. Each entry names the agent, the action, and the rule that decided, and leaves out request arguments and response bodies, so the log of your inbox doesn't become a second copy of it.

On Personal the log is the Activity tab, and it keeps seven days. Read it on a slow afternoon and you'll learn what your assistant thinks "tidy my inbox" means.

Calls run server-side, and your Gmail credentials never reach the agent or the model (security).

One default is worth changing. Policies fail open by default. Fail closed is an opt-in policy setting. For an inbox, it's the one to pick.

What none of this does is stop the model reading injected text. Nothing does. A policy limits what an injected instruction can make the assistant do next, which is the part that costs you.

Keep work and personal mail apart

Connect work and personal inboxes as separate accounts, and give each its own rules. They carry different risks and, on the work side, a different owner.

On a work account, your admin may have already decided. Claude's Microsoft 365 connector leaves work-account write actions to admins, and ChatGPT's Outlook connector needs an Entra admin's approval before it runs. Permissions an admin granted on your behalf are ones you can't revoke yourself. In Arc0, policies are set per agent, so a work assistant and a personal one can run different rules.

The work inbox is the one with the most to forward and the least of your say. Ask your admin what's granted before you add an assistant on top.

Review and revoke what you've already granted

Disconnect in the assistant, then check the provider's own page too. Claude's help spells out both steps, and the provider page is where you'll also find the grants you forgot you gave.

Google.

  1. Open myaccount.google.com/linkedapps.
  2. Select Access to your Google Account.
  3. Pick the app, then select See details.
  4. Select Remove access, then Confirm.

The app loses access, but anything it already copied stays with it.

Claude. Go to Customize, then Connectors, select the connector, choose Disconnect, and confirm. To remove it fully, open myaccount.google.com/connections, find Claude, and choose Delete all connections.

ChatGPT. Go to Settings, then Plugins, and choose the app or connected account; disconnecting Gmail also stops related mailbox monitoring.

Cursor. Toggle the MCP server off or remove it in Customize.

Microsoft personal account. As of November 2026 these labels come from one Microsoft support article, so they may drift. Go to accounts.microsoft.com, open the Privacy tile, then Overview, Other privacy settings, and Apps and services. Find Apps and services that can access your data and choose Remove these permissions.

Microsoft work or school account.

  1. Open myapplications.microsoft.com.
  2. Hover over the app and select the ... menu.
  3. Choose Manage your application, then Revoke Permissions.

Anything your admin granted stays until the admin removes it.

Arc0. In My apps, choose Disconnect. People using a product built on Arc0 get their own connected-apps page to see and revoke grants.

While you're on those pages, look for the app you connected for a single errand and forgot. It hasn't forgotten you.

Checklist before you click Allow

Shorter than the consent screen, and more candid about sending.

  • Read the consent screen and find the narrowest scope it asks for.
  • Allow search, read, and drafts.
  • Make send, reply, and forward ask you every time.
  • Deny delete, filters, forwarding, and settings, or don't expose them at all.
  • Keep work and personal inboxes on separate connections with separate rules.
  • Keep a log of every call, and actually read it.
  • Put a date in your calendar to review and revoke.

Arc0 is our pick because it puts one policy check in front of calls from any MCP client you connect, Claude, ChatGPT, and Cursor included. Get a free personal MCP URL from Arc0, connect Gmail, and add the URL to your client with the MCP client setup guide. Then go back and summarize this morning's threads. That's all you wanted in the first place.

Get started

One plug. Every app. Any agent.

Your users connect their apps once, under your brand. Your agent, or Claude, ChatGPT and Cursor, acts through one MCP endpoint, with every action on the record.