Should you let an AI agent into your email? Not your main inbox with full access, at least for now. This guide explains exactly what an agent can see once it is in, the one permission mechanic almost every guide gets wrong, what has actually gone wrong so far and how to limit, audit and remove access - checked against Google's, Microsoft's, OpenAI's and Anthropic's own documentation.
An AI agent is software that takes actions for you - our guide to what an AI agent is covers the background. Letting one into your email means granting a standing permission to read your messages, and perhaps to send or delete them too, and email is not just another app: password resets land there, other people's confidences live there and whoever can read it can usually become you.
Last checked: 5 November 2026, against the National Cyber Security Centre's and Information Commissioner's Office's published material, Google's, Microsoft's, OpenAI's and Anthropic's own documentation, the Financial Ombudsman and the researchers named below. None of this was tested on a UK account. We update this guide when things change.
Should you let an AI agent into your email?
Not your main inbox with full access, for three reasons. Email is the key to your other accounts - password resets land there. Hidden instructions can steer an agent - anyone can email you, an agent may treat the text as a command, and the NCSC says it is "very possible" this "may never be totally mitigated". Agents make mistakes with nobody attacking them - Meta's own help page tells Muse users to monitor its actions and intervene as needed.
If you want to try one anyway: use a spare account or the most limited connection the product offers, keep approval switched on for anything it sends and remove the access when you finish. The rest of this guide is the detail behind that sentence.
What "email access" actually means
Usually it means a connector: an approved link between the agent and your email provider, built on a standard called OAuth. You choose to connect Gmail or Outlook in the agent's app; your provider shows a consent screen listing what the app asks to do; you click Allow; the app receives a token - a limited pass that does not contain your password - and uses it whenever its design calls for it, watched or not.
Here is the mechanic almost every guide gets wrong: you generally cannot edit an app's requested permissions scope by scope. The developer defines the request; for the connection you are approving, you take the requested permissions as a set or decline. Depending on the service you may be able to connect fewer apps or accounts - OpenAI, for instance, documents choosing which Google apps to connect - but there is no menu at Google or Microsoft for trimming a request to "just this folder" or "read but never send", and your control afterwards is removal, not reduction. That makes the consent screen your most important decision point: it shows the access ceiling. One genuine mitigation exists where offered: Google's account permissions support time-limited access for some apps, with 30 and 180 days as its documented durations - not every Gmail connector offers it, but take it when it appears.
Never give an agent your email password, and be wary of app passwords - Google says they are not recommended and unnecessary in most cases. One correction to the folklore, though: Google app passwords are individually revocable, and all of them are revoked automatically when you change your Google password. The never-use-them advice stands because they grant broad access with none of a connector's structure - not because you cannot take them back. (One mainstream exception to OAuth-everywhere exists: ChatGPT's iCloud Mail connection uses an Apple app-specific password, per OpenAI's own pages.)
Connectors are not the only way in
Everything in this guide about permission screens and revocation applies to connectors. Other routes skip all of it. Logged-in browser sessions: a browser agent can reach Gmail through a session you are already signed into - no consent screen, no permission list, as the Comet demonstration below showed. Desktop apps: Meta's Muse Mac app reportedly reads local apps, including mail, with per-app access levels. Your password: Instinct's privacy policy says you may give its assistant your username and password for third-party accounts - which is exactly what you should never do. Removing access at Google or Microsoft does not touch any of these routes.
The screen shows the ceiling, not the plan
A permission screen tells you the most a connected app could do; a well-behaved app with permission to send may ask you before every send. Both halves are real: Anthropic's own consent screen for Gmail shows send rights, and Claude then asks approval before each send-class action by default. Judge an app by the ceiling anyway, because a bug or a hijacked agent can reach everything the grant allows.
What can an agent see? The Gmail and Outlook permission ladder
These are the main connector permissions - each row is a "scope", one named permission an app can request - with Google's own consent-screen wording where it has one. Google also classes its mail scopes as Sensitive or Restricted; the Restricted ones trigger extra vetting of the app.
| Permission | Service | What it allows |
|---|---|---|
| gmail.metadata | Gmail | Labels and headers - who wrote, when, usually the subject. Not the body |
| gmail.readonly | Gmail | "View your email messages and settings" - reads everything, changes nothing (Restricted) |
| gmail.send | Gmail | "Send email on your behalf" - sends as you, reads nothing (Sensitive) |
| gmail.compose | Gmail | "Manage drafts and send emails" - drafting includes sending (Restricted) |
| gmail.modify | Gmail | Read, compose and send, excluding permanent deletion (Restricted) |
| mail.google.com (full) | Gmail | "Read, compose, send, and permanently delete all your email from Gmail" |
| Mail.ReadBasic | Outlook | Mail details without body or attachments |
| Mail.Read | Outlook | Reads your mailbox |
| Mail.ReadWrite | Outlook | Creates, reads, updates and deletes mail - cannot send |
| Mail.Send | Outlook | Sends mail as you |
Three readings matter. Read-only still means everything: gmail.readonly covers old threads, password resets, bank notifications, medical letters and other people's messages - no scope limits an app to one sender or folder. Read-only does not stop a leak: it stops sending and changing, not what the agent read leaving by another route, as the incident table below shows. "Drafts" may include sending: Google's compose permission covers both. A summariser needs to read, not send or delete - if an app asks for more than its job requires, treat that as your answer.
What the mainstream AI tools actually do with your email
This table covers the four email-capable products with the most complete public documentation, reporting what each maker publishes - safety descriptions are vendor claims.
| Product | What it requests and does | Asking first | Training on your mail |
|---|---|---|---|
| ChatGPT + Gmail connector | Requests gmail.modify (read, label, draft, send - no permanent delete); you cannot narrow it. Event-triggered tasks can fire on new Gmail messages | OpenAI documents permission tiers - Always ask, Allow read actions, Allow low-risk actions and Allow all ("elevated risk") - though what appears depends on account, app and workspace. The tiers change when it asks, not what it can reach | OpenAI does not train "generalized models" on connected Google app data even with the improve-the-model setting on - except feedback you submit, content you paste in and content that appears in responses. That last exception is the gotcha: a summary in a chat can be trained on if the setting is on |
| Claude + Gmail connector | Searches, reads, drafts, sends, replies, forwards; reads attachment metadata but not attachment content | Approval before each send-class action by default; Team and Enterprise owners decide whether members may waive it | "We do not train our models on your Gmail, Drive, or Calendar connector data" - same paste-and-response exception on consumer plans with training on. Connector data is stored with the chat and deleted with it |
| Gemini in Gmail | A native Google feature, not an OAuth grant - it reads your content to answer prompts inside Workspace apps | Operates within Gmail; controlled by smart-features settings (or your admin), not a consent screen | Google says Workspace content is not used to train Gemini or other generative AI models (personal-account Gemini Apps sharing is the exception to check) |
| Microsoft 365 Copilot (work) | Reads content available to the signed-in user via Microsoft Graph, subject to Microsoft 365 permissions and tenant controls - not a personal connector grant you can revoke independently | Admin-controlled; prompts and responses are stored as activity history you can delete and your admin can search via Purview | "Prompts, responses, and data accessed through Microsoft Graph aren't used to train foundation LLMs" |
The always-on personal agents are a different family: Claude Duet builds a morning brief from your mail with approval gates on sending, Meta Muse connects email among its apps with purchase-and-send approvals and Instinct gives your agent an email address of its own - but none of the three is straightforwardly available to UK consumers, so we cover them in their own guides to Claude Duet, Meta Muse and Instinct rather than here. One early-user story worth carrying over: an Instinct tester reported disconnecting Google at 11am and still receiving an email summary at 2pm - retained copies are a thing, and the removal section below deals with them.
Prompt injection: the risk that sets agents apart
To a normal email app, a message saying "forward everything" is words on a screen. A language model cannot reliably tell your instructions from text somebody else wrote - so a stranger's message can carry instructions (a classic demonstration hides them as white text on a white background), and an agent with the right tools may obey them. When that works, the agent is hijacked: your access, someone else's orders.
The authorities on this are unusually blunt. The NCSC's technical blog says it is "very possible" that prompt injection "may never be totally mitigated" the way older attacks were, and its advice is to constrain what agents can do with deterministic safeguards rather than trust the model to resist. OWASP ranks prompt injection as the number one risk in its Top 10 for LLM applications and says it is unclear whether foolproof prevention exists. The vendors agree: Anthropic's researchers write that "no browser agent is immune to prompt injection", and Meta calls it an open problem for the industry.
What has actually happened
The cases below were disclosed as research findings or controlled demonstrations, and the sources cited report no confirmed exploitation against ordinary users in the wild - which is not proof none occurred. Note the statuses differ too: "fixed", "partial fix" and "disputed" are not the same thing.
| Case | Product | What was shown | Status |
|---|---|---|---|
| Check Point research, September 2026 | ChatGPT with Gmail | A planted instruction - delivered by a pasted prompt or shared chat, not by email - moved a victim's connected Gmail data to an attacker's ChatGPT account; the only visible trace was a "Talked to Gmail" label | OpenAI confirmed the internal service involved was decommissioned; no use in the wild reported |
| ClaudeBleed, public May 2026 | Claude in Chrome extension | Another extension with no special permissions could make Claude send email for an attacker and pass mail summaries out | Reported to Anthropic 27 April; fix shipped in v1.0.70 on 6 May; the researchers call the fix partial |
| ShadowLeak, public September 2025 | ChatGPT Deep Research with Gmail | Hidden email text made the agent exfiltrate data from OpenAI's own servers - zero clicks from the victim | Reported June 2025 by Radware, fixed in 2025, publicly confirmed 3 September; never seen exploited |
| Comet, disclosed August 2025 | Perplexity's Comet browser | Hidden text in a Reddit comment made the assistant open Gmail in the logged-in browser, read a one-time code and post it publicly - no connector, no consent screen involved | Disputed: Brave said not fully mitigated after a patch; Perplexity said fixed |
| Gemini summaries, July 2025 | Gemini for Workspace | Hidden email text made Gmail's AI summary display a fake security warning with an attacker's phone number | Google said it had seen no such incidents in the wild; mitigations were still rolling out |
| EchoLeak, CVE-2025-32711, June 2025 | Microsoft 365 Copilot (work product) | One crafted email, zero clicks, could make Copilot leak sensitive data to an attacker's server - rated Critical, CVSS 9.3, described as the first known zero-click attack on an AI agent | Fixed server-side by Microsoft with no customer action; never exploited |
Note what the table implies about read-only: in the Check Point and ShadowLeak cases the data left through the AI service itself, so a read-only grant would not have stopped either. And note Comet: browser agents in signed-in sessions bypass this entire permission model, which is why they should be kept away from your email altogether.
What goes wrong with no attacker involved
Plenty. The best-documented case: Tom's Hardware reported in February 2026 that an OpenClaw agent deleted more than 200 emails from the inbox of a Meta alignment director "despite repeated commands to stop" - she had asked it to get approval before acting, and reportedly had to terminate it at the machine. A small test inbox had reportedly gone fine first. No hacker involved; just an agent, a vague standing instruction and bulk access.
Three things raise the odds. Vague jobs: "tidy my inbox" leaves the agent to decide what tidy means; "list newsletters I have not opened" does not. Bulk work: software repeats one wrong decision across a mailbox until someone notices. Approval fatigue: Anthropic calls it "a serious issue" - asked to approve everything, you stop reading what you approve.
Why your inbox is the one to protect
The reason is structural: someone in your email can read private information including banking details, impersonate you and reset your other account passwords - which is why security guidance treats a compromised inbox as a compromised life. An agent is not a criminal, but it holds the same keys - and it can read every bank notification and one-time code that lands there.
Your bank and your refund rights. Never give a bank login to an AI tool - UK banks' terms say never share it with anyone (NatWest's fraud pages say exactly that). Sharing does not automatically cost you a refund: your bank must show gross negligence, a bar the Financial Ombudsman calls "very high" - but none of this has been tested on AI agents, and in the US, Bank of America has emailed customers - TechCrunch reported, with the bank confirming the email - saying they are responsible for activity resulting from credentials shared with an AI tool. If you want software to access financial information, use a properly authorised open banking provider rather than giving anything your bank credentials - and remember that is a payments safeguard, not a substitute for securing your email.
Other people's data. Your inbox holds what friends, family and colleagues told you, and they never agreed to an AI reading it. The ICO's agentic AI report notes that an agent's data may also include information about third parties the user interacts with. One reassurance for home users: for purely personal or household use, an individual is generally not subject to the 72-hour breach-notification duty that falls on organisations acting as controllers - though that position can change where the processing goes beyond the household context. And whatever the law says, the people in your inbox might reasonably expect you to tell them if something leaks - an ethical point rather than a legal one.
How to stay in control
The security principle is least privilege - a job gets only the access it needs - plus a human approving what matters. Applied to one person and one inbox:
Try a spare account first. A separate address holding nothing sensitive, signed up to a few newsletters, shows you how the agent behaves. Do not forward your main mail to it or use it as a recovery address. Read the consent screen as a verdict. You cannot trim the request, so the list is the decision: a summariser asking for send or delete rights has told you what you need to know. Keep browser agents away from signed-in email. Use a browser profile that is not signed in - permissions do not apply to that route at all. Keep approval on, at the strictest tier. In ChatGPT that means Always ask; in Claude, per-action send approval is already the default; avoid every "don't ask me again" option for email. Approval catches a forward you did not ask for - it cannot catch data leaving by a route that never asks. Check its work. Read Sent, Trash and any activity log for the first few runs - what it did, not what it said it did. A clean log is not proof. Diarise the removal. Put a date in your calendar to review or remove the access; standing grants are how six-month-old connections are forgotten.
When to say no
| The job | Our view | Why |
|---|---|---|
| Summarising newsletters in a spare account | Reasonable | Little to leak; mistakes cost little |
| Finding an old order in your main inbox | One-off only: connected, watched, removed | It sees everything while connected, so keep it short |
| Drafting replies you send yourself | Spare account yes; main inbox not yet | Drafting scopes can include sending |
| Sending without asking, or bulk deletion | No | A sent email does not come back, and bulk work repeats mistakes |
| A standing connection to an inbox full of money, health or legal mail | No | One leak costs more than the convenience is worth |
Red flags, whatever the product: it asks for deletion rights when its job is summarising; you cannot find an approval setting or an activity view; you cannot find what it keeps after disconnection or how to delete it; and no retention periods are published anywhere - true today of more than one major agent.
How to remove an agent's access to Gmail and Outlook
Remove access at your email provider and inside the agent's own settings - both, because each covers what the other misses. Type the addresses yourself; never follow a link from a message claiming an agent was compromised.
Google: go to myaccount.google.com/connections, choose the app, then Remove access and confirm. Google is explicit that removal does not delete data the app already holds - you may need to ask the developer to delete it. Microsoft personal accounts: account.live.com/consent/Manage, choose the app, then Remove these permissions. Microsoft work accounts: myapplications.microsoft.com, open the app's menu, Manage your application, Revoke Permissions - and note you can only revoke what you consented to yourself; admin-granted permissions are the admin's to remove. Inside the products: ChatGPT's connected apps live under Settings and its apps/plugins section - and note OpenAI's guidance that changing an action permission does not revoke the provider grant; disconnecting the app does. Claude's connectors live under Settings, Connectors. These paths move as products update, so if a menu is not where we say, search the provider's help for "remove third-party access" - the destination matters more than the route. Gemini in Gmail has no app grant to revoke on the Google connections page - it is a native feature rather than a third-party OAuth connection, governed by Gmail's smart-features settings and, on Workspace accounts, administrator controls.
Then delete what it kept: disconnect first, delete inside the product second, check it worked third. In ChatGPT, deleted chats normally purge within 30 days, but memory can retain connector-derived information and disconnecting does not delete existing chats or memories - clear those separately. In Claude, deleting the chat deletes the retrieved mail data stored with it.
What to do if your agent is hijacked
Signs: email you did not approve, contacts asking about odd messages, new forwarding rules or filters you did not create. Then, in order: stop it - cancel running tasks, remove access at the provider using the steps above, remove any browser extension or desktop app it uses; secure the account - change your email password and anything sharing it, sign out other sessions, turn on two-step verification; check forwarding rules and filters - attackers add these to keep copies flowing; check Sent and Trash - Gmail's Trash keeps deleted mail for 30 days, though full-scope access can bypass Trash entirely; tell people - anyone who received strange messages, and your employer if work mail is involved; check your bank - through official channels, at once, if anything financial was in reach; and save the evidence before asking the maker to delete your data.
Work email deserves its own sentence: do not connect it on your own initiative at all. Your employer controls that mailbox, enterprise AI tools are governed by tenant admins rather than your settings, and both ChatGPT and Claude offer businesses controls that block exactly this. Ask IT first, in writing.
The sensible first step: limited and easy to remove
The answer to the title is a conditional no: not your main inbox, not full access, not yet. The workable experiment is a spare account or the most limited grant available, approvals at the strictest setting, a diarised removal date and a check of what stayed behind. A clean trial tells you how an agent handles dull mail - it does not tell you how it handles a hidden instruction or a bigger inbox, so treat every step up as a new decision. The agents are improving and the documentation is genuinely better than a year ago; your inbox is not getting any less valuable, so let the makers earn it last.
Frequently Asked Questions About AI Agents and Email
Are AI agents safe to use with email?
Not risk-free. Researchers have repeatedly demonstrated hidden instructions steering agents with email access, and agents have made costly mistakes unprompted. No attack on ordinary consumers has been confirmed in the wild, which is not proof of none. Start with a spare account, keep approvals on and remove access when you finish.
What can an AI agent see in my email?
Whatever its granted scope allows - and the main read permissions on Gmail and Outlook cover your whole mailbox, not one sender or folder. Metadata-only scopes see headers without bodies. A browser agent in a signed-in session is not limited by any of this.
Can I limit which permissions an AI agent gets?
No - this is the key mechanic. The developer decides what the app requests, and you grant the whole list or refuse. Your controls are the consent screen (read it as a verdict), any time-limited sharing Google offers, the agent's own approval settings and removal afterwards.
Does ChatGPT or Claude train on my Gmail?
Both say no for connector data: OpenAI does not train "generalized models" on connected Google app data, and Anthropic says it does not train on Gmail, Drive or Calendar connector data. Both carry the same exception - content you paste into a chat, or that appears in the AI's responses, can be used if your training setting is on. Check that setting.
Is prompt injection preventable?
Not reliably, on current evidence. The NCSC says it is very possible it may never be totally mitigated; OWASP ranks it the top LLM risk and says foolproof prevention is unclear; Anthropic says no browser agent is immune. Narrow access, approval gates and audit trails limit the damage rather than prevent the attack.
How do I remove an AI agent's access to my Gmail?
Go to myaccount.google.com/connections, select the app, then Remove access - and also disconnect inside the agent's own settings. Removal stops new access but does not delete what the app already holds, so delete that inside the product too.
How do I know if my AI agent was hijacked?
Look for messages you did not approve, contacts reporting odd emails and forwarding rules or filters you did not create. You may see very little - in one demonstration the only trace was a small label. Remove access first, investigate second.
Can my employer's Copilot read my work email?
Microsoft 365 Copilot reads only what you already have permission to see, is governed by your employer's admins rather than your settings, and Microsoft says the data is not used to train its foundation models. Your prompts and responses are stored as activity history your admin can search - so treat work Copilot as work.
Should I connect an always-on agent like Claude Duet or Meta Muse to my email?
UK consumers cannot yet - neither is available here, and Instinct is an invite-only beta. When they arrive, the same tests apply: what scope does it take, what does it ask before sending, what stays after you disconnect and where is the activity log.