Summary

For business owners connecting AI to company tools, this guide explains how scopes, account rights and resource assignments set the agent’s reach. It gives concrete permission comparisons and a repeatable access review.

What you'll learn

  • How Gmail separates mailbox reading, drafts and sending
  • How Drive file-level access differs from an account-wide grant
  • Why Slack token type and channel membership matter
  • How delegated, application and selected Microsoft permissions differ
  • How to review, revoke and separately check retained data

Permission scopes decide your exposure. Not the model, not the vendor's security page — the specific strings you approved when you clicked Allow.

Two scopes can sit one line apart on a consent screen, read almost the same, and differ by everything: one document against every file in the account.

The worry behind this is common and measured. Among EU enterprises in Eurostat’s covered non-financial sectors, excluding agriculture, forestry and fishing and mining and quarrying, with 10 or more employees that had ever considered using AI, 52.72% cited concerns about violation of data protection and privacy in Eurostat’s 2025 data.[8] The first check is in your account’s access settings: which application is connected, which permissions it holds and whose account it uses.

For the underlying distinction between a model, an agent, and a workflow, start with what an AI agent is.

What can an agent actually see?

Three boundaries set an agent's reach:

  • Scope — which permissions you granted.
  • Identity — which account the agent acts as.
  • Membership — which files, channels, or spaces it was explicitly added to.

Gmail: reading and sending are separate

As of October 2026, Google separates mailbox reading, draft management and sending.[1] The short Gmail and Drive names below normally have the full prefix https://www.googleapis.com/auth/; https://mail.google.com/ is a complete scope by itself.

Gmail scopeWhat it grantsCan read mailCan send
gmail.metadataLabels and headers, not message bodiesPartialNo
gmail.readonlyView messages and settingsYesNo
gmail.labelsSee and edit labelsNoNo
gmail.sendSend email on your behalfNoYes
gmail.composeManage drafts and sendDrafts only, not all received mailYes
gmail.modifyRead, compose, send; no permanent deleteYesYes
https://mail.google.com/Full access including permanent deleteYesYes

The read-and-send split is the useful part. An outreach workflow that reads with gmail.readonly and sends with gmail.send has two narrow permissions instead of one broad one. An agent holding gmail.modify instead has both capabilities plus the ability to modify anything in the mailbox.

Google Drive: one file or the whole account

As of October 2026, Google Drive distinguishes file-specific access from account-wide grants.[2]

Drive scopeWhat it grants
drive.fileCreate files; read or modify files created by the app, opened with it or shared through its file picker.[2]
drive.metadata.readonlyFile metadata only, no content
drive.readonlyView and download files available to the connected account, subject to file restrictions
driveFull read and write across the entire Drive

An agent that needs one spreadsheet should hold drive.file. With drive.readonly, files available to that account can include contracts, payroll and shared documents. Check the account’s actual rights and any download restrictions, rather than assuming the scope bypasses them.[2]

One task, two permissions

Narrow access

drive.file
01

The agent can work with files it created or that you opened or shared through its file picker.

Account-wide access

drive.readonly
02

The agent can view files available to the connected account; file restrictions still apply.

The task is identical.

The exposure is not.

Slack: scope and channel membership both matter

As of October 2026, a Slack bot’s channel membership adds a second boundary. Its history scopes are limited by channel membership as well as by scope: channels:history grants messages in public channels "that your Slack app has been added to." The scope sets the ceiling; being added to a channel sets the bot’s reach. User tokens have different history rules, including public-channel reach, so check the token type too.[12]

Slack scopeGrantsClass
channels:readBasic public-channel informationRead
channels:historyMessages in channels the app was added toRead
im:historyDirect messages the app was added toRead
files:readFiles in conversations the app was added toRead
users:readPeople in the workspaceRead
chat:writePost messages as the appWrite

Microsoft 365: Outlook, OneDrive, SharePoint and Teams

Microsoft 365 tools ask for Microsoft Graph permissions; many have delegated and application forms, while some offer only one. A delegated permission acts as the signed-in user and reaches only what that user can reach. An application permission runs without a signed-in user, and Microsoft's own descriptions say it covers "all mailboxes" or "all site collections". These are the relevant grants as of October 2026. Check whether the application uses delegated or application access before comparing names.[9]

Graph permissionDelegated: what it grantsApplication versionCan read content
Mail.ReadBasicThe user's mail properties except body, preview, attachments and extended propertiesSame, in all mailboxesPartial
Mail.ReadThe signed-in user's mailboxAll mailboxesYes
Mail.SendSend mail as the userSend as any userNo
Mail.ReadWriteCreate, read, update, delete mail; does not include sendingAll mailboxesYes
Files.ReadThe signed-in user’s files; personal accounts also include shared filesNot offeredYes
Files.Read.AllEvery file the signed-in user can accessAll files in all site collectionsYes
Sites.SelectedExplicitly assigned sites, intersected with the signed-in user’s rightsExplicitly assigned sitesAssigned sites only; consent alone gives no access
ChannelMessage.Read.AllTeams channel messages, on the user's behalfAll channel messagesYes

Two details decide the real reach. Application permissions need an administrator's consent, and Microsoft notes that administrators can configure an application access policy to limit an app holding Mail.Read or Mail.Send to specific mailboxes instead of the whole organisation. For files, Sites.Selected needs both consent and an explicit resource assignment; delegated use is also limited by the signed-in user’s rights. It selects sites rather than individual Drive files, so the two scopes are not interchangeable.[13] Exchange also offers application RBAC for scoped mailbox access; check all effective grants, since a broad parallel grant can preserve broad access.[14] Microsoft Copilot follows the same rule from the other side, reading whatever each user can already open; our Copilot GDPR checklist covers what that means before switching it on.

Why do permissions matter more than the model?

A better model does not narrow an agent’s permissions. Review its identity, scopes and available actions independently of which AI you choose; OWASP describes how excessive inherited rights enable unauthorised actions.[4]

Choose the permissions for the job, not for the model’s apparent ability.

An access problem can arise from a permission related to the task that reaches more files than the task needs.

Start with a concrete file task: a reporting assistant needs a spreadsheet you select. Google’s drive.file lets an app work with files created by it or explicitly opened/shared through the picker. A whole-Drive reading grant reaches the connected account’s available files instead.[2]

Test that distinction with a harmless document outside the agreed scope. Ask the operator to show that the narrow setup cannot retrieve it. Then check the intended spreadsheet still works. A connection that passes only the successful-task test has not yet shown you its boundary.

If the proposed connector requires broad access, ask why: does it search across the account, or does it merely need the selected document? Agree the account, files and actions before approving the grant. A dedicated account may make that inventory easier, but its own shared folders and permissions still need review.

Agent systems can inherit privileges they do not need

OWASP lists Identity and Privilege Abuse as ASI03 in its OWASP Top 10 for Agentic Applications 2026. The 2026 report was published in December 2025.[4]

Its diagnosis is precise:

Without a distinct, governed identity of its own, an agent operates in an attribution gap that makes enforcing true least privilege impossible.

OWASP also names the mechanism that produces over-provisioning in multi-step systems: un-scoped privilege inheritance, where a high-privilege component delegates a task "passing its full access context," so "a narrow worker then receives excessive rights."

Older guides still use the previous name, Excessive Agency. OWASP says ASI03 "is the agentic evolution of Excessive Agency (LLM06:2025)."

Use the taxonomy appropriate to the system you are assessing: an agentic workflow’s inherited access needs review alongside its tool actions.[4]

Read-only is narrower, not automatically safe

Read-only tool scopes remove write/send operations for that connection, when those are genuinely the only effective grants.[1][2] The agent can still be influenced by what it reads.

Google Cloud's Office of the CISO describes the route directly. Agents are "vulnerable to 'jailbreaks,' which are clever prompts that are designed to bypass their safety features and expose confidential information."

Its worked example is an agent tasked with summarising email, then "fed a jailbreak prompt that causes it to leak confidential information to an unauthorized recipient."

Read access still accepts untrusted input

The content an agent reads is untrusted input. An email, a shared document, or a calendar invite can carry instructions aimed at the agent rather than at you. Read scope determines what the attacker's instruction can reach; it does not stop the instruction arriving.

This is why scoping and gating are two controls, not one. Narrow scopes limit the blast radius. A gate on outbound actions limits what an influenced agent can do with what it found.

Least privilege applies to agents the same way it applies to people: the agent gets the bare minimum access to do its job, and nothing more. A tool connection without a revocable, logged permission scope is an open door — and read-only access is a smaller door, not a closed one.

What three controls decide exposure?

Google Cloud's Office of the CISO frames agent security on three traditional controls. Anton Chuvakin, Security Advisor in that office, sets them out as authentication, authorization, and auditability.

ControlWhat it means for an agentWhat it looks like in practice
Authentication"every agent has a unique identity, allowing you to verify that it is what it claims to be"A dedicated account for the agent, not a person's login
Authorization"only give an agent the bare minimum access it needs to perform its job and nothing more"drive.file over drive.readonly; gmail.send over gmail.modify
Auditability"logging every action the agent takes… a clear record, or audit trail"A per-action log naming the tool, the data, and the outcome

Treat identity as a separate design decision. An agent acting through a person’s grant reaches the intersection of the granted scopes and that person’s permissions. Check both before connecting it.[9]

It also becomes hard to trace because every action in the log belongs to the human. A dedicated account helps separate responsibility; still check its grants and keep application-specific action logs.

The same task shown twice: on the left an agent with drive.file reaching one document, on the right the same agent with drive.readonly reaching files available to that account.
Same job, two scopes. The scope decides the exposure, not the workload.

Where an agent's permission gate sits

The gate is a separate decision from what the agent may reach. In Yardwork implementations, read access is the default. Sends and other high-impact actions sit behind an approval enforced in code rather than in a prompt.

The agent holds the send capability, but nothing leaves until a person reviews and approves it. That describes implementation practice scoped per engagement. It is not a security guarantee, an audit, or a certification.

Why a prompt is not a permission boundary

A prompt is not a permission boundary. An instruction telling a model not to send is a preference it may fail to honour. A gate in the sending path can require approval before release. Test every available sending route and who can change the gate; the condition works only when the workflow cannot use a different route.

Approval workflows show how to hold, review and release an action. AI agent governance covers the wider control design.

What stays behind after the task ends?

An agent's permissions describe what it can reach. They say nothing about what it keeps.

Spain's data protection authority, the AEPD, addressed this directly in Agentic Artificial Intelligence from the Perspective of Data Protection (V1.1, February 2026). The agent's memory stores personal data. Its logs can reach further:

The logs will store information about both the users of the agentic AI and the people subject to the processing, it could even store data on people who should not be data subject.

Three places hold data after a run finishes:

  1. Working context — the material the agent pulled in to do the task.
  2. Long-term memory — summaries, profiles, and embeddings that persist so the agent need not be re-briefed.
  3. Logs — the audit trail, which is itself a store of personal data.

The AEPD recommends separate retention periods for each category rather than one global rule. It also recommends allowing persistent memory to be disabled when the processing does not need it.

The practical question: which of the three exists in your deployment, and how long is each kept?

"We do not delete anything" is a common and honest answer. It is also a decision, whether or not anyone made it deliberately.

Data also leaves. Model providers, enrichment APIs, translation and summarisation services each receive whatever the agent sends them. If your business falls under the GDPR, those recipients and any transfers outside the EU carry their own obligations — the GDPR and AI agents guide maps which article each action engages.

How do you review and revoke access?

Reviewing and revoking agent access is the part you can do today, without asking anyone.

1

Inventory the connections

Open your account's third-party access settings. In a Google Account this is the third-party apps and services section; in Slack it is the workspace's installed apps list. Every connected app is listed with the scopes it holds.
2

Read the scopes, not the app name

A tool called "meeting notes" may hold drive.readonly. Compare each granted scope against what the tool needs to do its job.
3

Ask which account it acts as

Record the account and the granted scopes together. A dedicated account helps separate the work, but review its own permissions and action logging as well.
4

Find the gate

Establish which actions run unattended and which require a person. If the answer is "the prompt tells it not to," there is no gate.
5

Revoke what is stale

Remove connections for tools no longer in use. Revoking access stops future reach; it does not delete data already held, so ask about that separately.
6

Set a re-review date

Permissions drift as workflows change. A connection granted for one task tends to outlive it.

Download the AI agent permissions checklist (English PDF).

A six-stage cycle: inventory, read scopes, check identity, find the gate, revoke stale connections, and re-review on a date.
Access review is a repeatable loop, not a one-time audit before launch.

What questions should you ask before connecting a tool?

The useful questions have answers you can check afterwards.

  • Which exact scopes does this integration request? Ask for the scope string, not whether it "accesses Drive." If the vendor cannot tell you, that is itself an answer.
  • Can it work with a narrower scope? drive.file is often enough for a tool that needs only the documents you hand it.
  • What identity does it act as? Look for a dedicated account rather than someone's personal login.
  • What is logged, and can I read it? A log you cannot inspect does not give you auditability.
  • Which third parties receive the data? Include model providers, enrichment services, and anything the agent calls mid-task.
  • What is retained, and for how long? Working context, long-term memory, and logs are three separate answers.
  • How do I revoke it, and what happens to data already held? Revocation and deletion are different operations.

Frequently asked questions

Can an AI agent see everything in my email? Only if you granted a scope that allows it. Gmail's gmail.readonly scope permits viewing messages and settings. The gmail.metadata scope permits labels and headers but not message bodies. The gmail.send scope permits sending and cannot read the mailbox at all. What the agent sees is set by which of these you approved, not by what the agent is capable of.

What is the difference between read and write access for an AI agent? Read access lets the agent retrieve information. Write access lets it change or create things, and send access lets it act outside your organisation. These are separate permissions in most tools. Gmail treats sending as its own scope, so an agent can be permitted to send without being permitted to read, or the reverse.

How do I check what permissions I gave an AI tool? Open your account's third-party access settings — in Google this is the third-party apps section of your Google Account, and in Slack it is the installed apps list for the workspace. Each connected app lists the scopes it was granted. Compare that list against what the tool needs to do its job.

What is least privilege for an AI agent? Least privilege means the agent holds the narrowest permission that still lets it finish the task, and nothing beyond it. Google Cloud's Office of the CISO defines it as giving an agent the bare minimum access it needs and nothing more. In practice it means choosing drive.file over drive.readonly when the agent only needs one document.

Is this page a security guarantee? No. This page explains how permission scopes work and how to inspect your own. It is not an audit, a certification, or a guarantee about any product or deployment, including Yardwork's. Scope definitions also change over time, so verify current documentation before relying on any specific scope behaviour.

Notes

  1. Gmail API — Choose Auth Scopes, Google Workspace developer documentation, retrieved October 14, 2026. Source for all Gmail scope strings and their stated grants.

  2. Google Drive API — API-specific authorization, retrieved October 14, 2026. Source for all Drive scope strings and their stated grants.

  3. Slack — Scopes reference, retrieved October 14, 2026. Source for Slack bot token scopes and the channel-membership qualifier.

  4. OWASP Top 10 for Agentic Applications 2026, published December 2025, reviewed within the OWASP project; full PDF retrieved October 14, 2026. Source for ASI03 Identity and Privilege Abuse, un-scoped privilege inheritance, and the relationship to Excessive Agency (LLM06:2025).

  5. Google Cloud Office of the CISO, "AI agent security: How to protect digital sidekicks (and your business)", Anton Chuvakin, September 13, 2025, retrieved October 14, 2026. Source for authentication, authorization, and auditability, and the email-jailbreak example. Vendor publication, attributed.

  6. AEPD, Agentic Artificial Intelligence from the Perspective of Data Protection, V1.1, February 2026, retrieved October 14, 2026. Source for the memory-and-logs passage and the retention recommendations. AEPD guidance is authoritative in Spain and persuasive, not binding, elsewhere in the EU.

  7. The file-task and boundary test are an implementation example of the permissions in source 2; no marketplace frequency is asserted.

  8. Eurostat, dataset isoc_eb_ai, retrieved via the Eurostat dissemination API on October 14, 2026; dataset last updated June 15, 2026. Indicator E_AI_BCDP, unit PC_ENT_AI_EC, geo=EU27_2020, size_emp=GE10, nace_r2=C10-S951_X_K, time=2025: 52.72%. The denominator is enterprises that had ever considered using any measured AI technology, and the dataset measures a range of AI technologies rather than agents.

  9. Microsoft Graph permissions reference, Microsoft Learn, page dated September 14, 2026, retrieved October 14, 2026. Source for all Microsoft Graph permission names, their delegated and application descriptions, admin consent, and the application access policy note.

  10. The access-review loop, the scope comparisons, and the provider question set in this guide are Yardwork work-product. Scope definitions change; verify current provider documentation before relying on any specific behaviour. This guide is not a security assessment, audit, or certification.

  11. Gmail draft retrieval, accessed October 14, 2026.

  12. Slack history by token type, accessed October 14, 2026.

  13. Selected permissions, accessed October 14, 2026.

  14. Exchange application RBAC, accessed October 14, 2026.