Start an AI agent on read-only email access. It can summarise, search and draft replies without being able to send anything, which removes the most direct way for a mistake or a hidden instruction to leak your data. Add send access only when you need it, and only with an approval step that shows you the recipient and the content. Calendars follow the same logic: viewing is low risk, but creating events with guests is a form of sending. This guide covers the choices for Meta Muse, OpenAI dots, Claude Duet and Grok Bot, based on official docs as of November 2026.
Why does the email scope matter so much?
Email is where most of your digital life ends up: receipts, bank alerts, medical letters, travel plans, password resets. It is also where strangers can put text in front of your agent. Anyone can email you, and an email can contain instructions written for the agent rather than for you.
That combination is two legs of Simon Willison's lethal trifecta: private data and untrusted content. The third leg is a way to send data out. Send access is the most direct version of that third leg. Read our lethal trifecta explainer for the full picture.
- Stranger emails you
- Agent reads hidden instruction
- Agent tries to forward data
- Read-only scope blocks the send
What are the scope levels?
Vendors use different names, but the levels are roughly:
| Level | What the agent can do | Risk |
|---|---|---|
| None | Nothing in this account | None |
| Read | Search, read, summarise | Sees everything; cannot send |
| Read + draft | Write drafts you send yourself | Low; you check before sending |
| Send with approval | Sends after you confirm | Medium; depends on reading prompts |
| Send without asking | Sends on its own | High |
| Delete or manage | Archive, delete, change filters | High; can hide its own actions |
General scope levels; exact options vary by agent and by Google or Microsoft connector, as of November 2026.
The last row deserves a note. An agent that can delete mail or create filters can remove evidence of what it did, so even a well-meaning mistake becomes harder to spot.
What does each agent offer?
Meta Muse
Meta says you choose access levels per connector, for example reading email without sending it, and can disconnect at any time (Meta launch post). Its security post adds two useful details: the email connector filters out one-time codes, password-reset links and magic login links, and processes that have read your data lose auto-allow for sending ("tainted egress"). The Mac Messages connector is separate, opt-in and needs Full Disk Access. See Meta Muse safety settings.
OpenAI dots
dots connects to email through apps and Custom Rules decide whether sending is automatic, pre-approved, asked or handed off. OpenAI says an auto-review step checks planned actions, such as an email's recipient and content, against your instructions and rules, and a dot cannot turn it off. Proactive research is read-only and enforced in code, per eesel's summary. Set sending to "ask before acting" to start. See dots Custom Rules setup.
Claude Duet
Duet uses the connectors you have in Claude, and switching an app off in Claude also switches it off in Duet. Before sending anything, Duet shows the action, every recipient and the content, then waits, according to Anthropic's safety article. Microsoft 365 was not yet available as of late October 2026. See Claude Duet approval settings.
Grok Bot
Each Bot has its own address at @mail.grokbot.com, and xAI's FAQ says unrequested outbound mail waits as a draft. Its "Ask first" rules cover sending, publishing, deleting and purchasing. If you connect your own email account, keep sending under "Ask first".
How should I set up calendar access?
Calendars look harmless but have their own risks:
- Viewing reveals where you will be and when, including home and travel details. Low risk to grant, but the data is sensitive.
- Creating events for yourself is reversible and low risk.
- Inviting others sends email from your account with your details. Treat it as send access.
- Accepting or declining on your behalf commits you to things. Keep it on approval.
- Deleting events can cause missed appointments. Keep it manual.
- Start with readLet the agent view your calendar to find free slots.
- Allow holds for yourselfLet it add events with no guests.
- Approve invitesRequire approval for any event with other people.
- Keep deletes manualDo not grant delete unless a task truly needs it.
When is send access worth it?
Send access saves time when you send many similar, low-stakes messages: confirming appointments, replying to routine requests, following up on orders. It is less worthwhile, and riskier, for anything involving money, personal details or people you do not know.
Pros
- Saves copying drafts into your email client
- Approval screens on Duet and dots show recipient and content
- Grok Bot's own address keeps unrequested mail as drafts
Cons
- One misread prompt can send private data to the wrong person
- Hidden instructions in incoming mail target exactly this permission
- Approval fatigue tempts you to stop reading prompts
How do I narrow access I already granted?
Change the level in the agent's connector settings if it offers one. If not, remove the agent's access in your Google or Microsoft account and reconnect with fewer permissions. Our step-by-step guide to revoking AI agent access covers both. For the broader principle, see least privilege for AI agents.
For a comparison of which agent handles email and calendar best overall, see our sister site's best AI agent for email and calendar. If your agent also reads texts, the same injection risks apply: see prompt injection via messages.



