Hot AI News

OAuth scopes: How an app gets read-only access to your data without your password

When Gemini reads your spending or a chatbot links to your bank, a permission system decides what it can see and do. OAuth scopes are the standard way to draw that line.

OAuth scopes: How an app gets read-only access to your data without your password
Photo illustration: Aiholics

OAuth is a standard that lets one app act on your behalf inside another service without ever learning your password. A scope is the label on that permission: it names what the app may see or do, such as “read calendar” but not “edit calendar.” It is the mechanism behind most “connect this app” buttons, including those AI assistants now use.

Google’s Gemini is a current example. As covered in our report on the Wallet connection, Google says Gemini can only read information from connected accounts and cannot move, manage or invest money. That kind of limit is what scopes are designed to express.

A hotel key card instead of your house key

Imagine you hire a cleaner. You could hand over your own house key, which opens every door at any hour. Or the building could issue a key card that opens only the kitchen, only on Tuesdays, and that you can cancel from your phone.

OAuth is the key card system. Without it, a budgeting app would ask for your bank username and password and could do anything you can. With it, the app receives a limited token instead, and your password never leaves the bank or Google.

Advertisements

Four roles and one token

A hand touches a permission toggle on an inset touchscreen showing four connected panels and a violet access token.
Photo illustration: Aiholics

The standard, OAuth 2.0, was published as RFC 6749 in October 2012 and edited by Dick Hardt. It names four roles:

  • Resource owner: you, the person who owns the data.
  • Client: the app asking for access, such as an assistant.
  • Authorization server: the service that checks who you are and issues permission.
  • Resource server: the system that holds the data and answers requests.

In the common flow, the app sends you to the authorization server with a list of scopes it wants. You sign in there, see a consent screen, and approve or refuse. The server then hands the app a short-lived code, which the app swaps for an access token. RFC 6749 describes that token as a string denoting a specific scope, lifetime and other access attributes.

From then on the app shows the token instead of a password. The data service checks what the token allows and answers only within that limit.

What a scope actually says

Per RFC 6749, a scope is a space-separated list of strings, and the authorization server decides what they mean. The server may also grant less than was requested, and if so it must say what it granted. Google’s OAuth documentation makes the same point: apps should check which scopes were actually granted.

This is where “read-only” comes from. A service can offer one scope that lets an app read transactions and a different one that lets it initiate payments. An app holding only the first cannot do the second, no matter what its code tries. Google’s documentation also says a token is valid only for the operations and resources in its scope, so a Calendar token does not open Contacts.

Google’s guidance on least privilege advises requesting scopes only when a feature needs them, rather than everything up front. It also says to disable features that cannot work without the related access.

Advertisements

Tokens expire and you can cancel them

Access tokens have limited lifetimes. To avoid asking you to approve again every hour, services can issue a refresh token, which only the authorization server sees and which obtains new access tokens. Google’s documentation lists several reasons these can stop working:

  • The user revokes the app’s access.
  • The token goes unused for six months.
  • The user changes their password while the app holds Gmail scopes.
  • An administrator restricts the service.

Revocation is the practical safety valve. Turning off a connection is meant to cut off the token, not just hide a setting. Google says the Personal Intelligence connections can be switched on or off at any time, though its announcements do not describe the underlying mechanism.

Where bank links and AI assistants fit in

Aggregators such as Plaid sit between apps and banks. According to Plaid’s documentation, its Link component lets a user pick a bank, log in, authenticate and choose which accounts to share. For banks that support OAuth, Link redirects the user to the bank and back. The app’s server then exchanges a temporary public_token for a permanent access_token and uses that for later requests.

Google says users can link credit card and bank accounts to Wallet through Plaid, and that Gemini can read but not move money. The announcement does not say how the read-only limit is enforced technically. OAuth-style scopes are the standard way to do it, but that is the general pattern and not a confirmed detail of Google’s system.

For AI assistants the stakes differ from a budgeting app. A chatbot reads the data as text and may repeat or summarize it in a chat. A scope controls what the assistant can fetch. It does not control what the model does with it afterward, how long it keeps it, or whether it gets a transaction wrong.

Advertisements

Limits of the scope model

Scopes are coarse. A scope like “read transactions” usually covers every transaction in the account, not just last quarter’s holiday spending. Users often cannot narrow it further, and consent screens are easy to click through.

The protocol has also had security problems. RFC 9700, the OAuth 2.0 Security Best Current Practice published in January 2025 by Torsten Lodderstedt, John Bradley, Andrey Labunets and Daniel Fett, collects them. It recommends:

  • Using PKCE (Proof Key for Code Exchange), which stops a stolen authorization code from being redeemed by an attacker.
  • Matching redirect addresses exactly, since pattern matching has led to real attacks.
  • Binding tokens to the app that received them, so a stolen token is less useful.

It also deprecates older, weaker modes of operation. Its advice is expected to feed into OAuth 2.1.

Finally, a read-only scope says nothing about retention. Google’s Wallet post does not state how long transaction data stays available to Gemini or whether chats using it improve models. Google’s Personal Intelligence announcement says Gemini does not train directly on your Gmail inbox or Photos library, but it trains on limited data such as prompts and responses after filtering or obfuscating personal data. Whether that covers Wallet data is not stated in the pages we read.

What to watch as assistants connect to more of your life

Three things are worth checking whenever an assistant offers a new connection:

  • What the consent screen lists. Look for read-only wording versus anything that can act.
  • Where to revoke. A trustworthy connection has a clear off switch, and the third-party service, such as the bank or Google account page, often has its own.
  • What happens to the data after reading. Retention and training terms are separate from scopes.

Rivals are moving the same way. Reporting on OpenAI’s finance tools in ChatGPT also describes bank connections through Plaid with read-only access. As more assistants gain such links, finer-grained scopes, shorter token lifetimes and clearer revocation will matter more than the headline promise that an assistant cannot spend your money.

Sign Up for the Daily AI Pulse

One email a day. All the stories that matter.

Daniel ReedAI Research, Safety & Ethics Analyst

Daniel Reed currently works as an AI Research, Safety & Ethics Analyst at Aiholics, writing about how changes in artificial intelligence are affecting and will affect scholarship,…

Leave a Reply

Your email address will not be published. Required fields are marked *

You may also like