# What is OAuth?

> OAuth is an IETF standard that lets an application get limited access to a web service, either for a person who approves it or on its own behalf. The application uses an access token, separate from the person's credentials.

Explainer by Boston Identity, 2026-10-25. 7 min read. https://insights.bostonidentity.com/glossary/oauth

- What it is: standard
- Defined by: IETF, RFC 6749 (2012)
- Who it's for: Anyone approving apps or agents that reach company data

## What is OAuth?

OAuth 2.0 is a framework the Internet Engineering Task Force (IETF) published in 2012 as RFC 6749, a Standards Track document [S1]. It lets a third-party application get limited access to a web service, either for a person who approves it or on its own behalf [S1].

Before OAuth, a person who wanted a printing service to reach their photos had to give it their password. The printing service then stored that password and gained access to everything behind it. The only way to cut it off was to change the password, which cut off every other app as well [S1 §1]. OAuth gives the application its own credential instead: an access token, a string that stands for a specific scope, lifetime and other access attributes [S1 §1]. An authorization server issues the token with the person's approval, and the application presents it to the service that holds the data [S1 §1].

The standard names four roles [S1 §1.1]. The resource owner can grant access; when it is a person, the RFC calls them the end-user. The resource server hosts the protected data. The client is the application asking for access. The authorization server issues tokens after it authenticates the resource owner and gets their approval. A single authorization server may issue tokens that many resource servers accept [S1 §1.1].

Think of it like this: A hotel key card is the closest everyday match. The front desk checks who you are once and gives you a card for one room until checkout, and the door never sees your passport. The comparison holds at theft too: like an OAuth bearer token, the card works for anyone who holds it [BI-FIND-2012-0001](https://insights.bostonidentity.com/ref/BI-FIND-2012-0001).

Not the same as: OAuth decides what an application may reach, not who the user is. Asked what they build today, one integrator named OpenID Connect for sign-in and OAuth for authorization [BI-FIND-2026-0484](https://insights.bostonidentity.com/ref/BI-FIND-2026-0484), and RFC 6749 itself warns that a sign-in built on its implicit grant needs extra checks [BI-FIND-2012-0005](https://insights.bostonidentity.com/ref/BI-FIND-2012-0005).

## How does OAuth work?

The RFC defines four ways to get a token, called grants, and prefers the one where the authorization server stands between the client and the person [S1 §1.2]. In that authorization code grant, the person's web browser carries messages between the client app and the authorization server, and a resource server holds the data. A client acting for itself can instead use the client credentials grant, where it gets a token with its own credentials alone [S1 §1.3.4].

[Figure: The person signs in only at the authorization server, so the client app ends up holding a short-lived code and then a token, never the person's password.]

1. **The client sends the person to the authorization server.** The client redirects the person's browser to the authorization server's authorization endpoint. The request carries the client's identifier, the scope it wants, a state value the client keeps, and a redirection URI, the address the server will send the browser back to [S1 §4.1]. The RFC says the state value should be used to prevent cross-site request forgery [S1 §4.1.1].
2. **The authorization server signs the person in and asks.** The authorization server authenticates the person through the browser and learns whether they grant or deny the client's request [S1 §4.1]. Because the person signs in only at the authorization server, their credentials are never shared with the client [S1 §1.3.1].
3. **The authorization server sends back a code.** If the person agrees, the authorization server redirects the browser to the client's redirection URI with an authorization code and the client's state value [S1 §4.1]. The code must expire shortly after it is issued, and the RFC recommends a lifetime of no more than 10 minutes [S1 §4.1.2]. The code is bound to the client identifier and the redirection URI [S1 §4.1.2].
4. **The client trades the code at the token endpoint.** The client sends the code to the authorization server's token endpoint, along with the redirection URI it used before [S1 §4.1]. The server must require client authentication from a confidential client, one able to keep its credentials secret [S1 §2.1] [S1 §4.1.3]. The client must not use a code more than once [S1 §4.1.2].
5. **The authorization server checks and issues tokens.** The authorization server checks that the code is valid, that it was issued to this client, and that the redirection URI matches [S1 §4.1.3]. If all holds, it returns an access token and, optionally, a refresh token, which the client can later present to get a new access token when the current one expires [S1 §1.5]. If a code is used twice, the server must deny the request and should revoke the tokens already issued from it [S1 §4.1.2].
6. **The client calls the resource server with the token.** The client asks the resource server for the protected data and presents the access token [S1 §1.2]. The resource server validates the token and, if it is valid, serves the request [S1 §1.2]. The RFC leaves out of scope how the authorization server and resource server talk to each other [S1 §1.1], and how the resource server validates a token [BI-FIND-2012-0007](https://insights.bostonidentity.com/ref/BI-FIND-2012-0007).

## Why does OAuth matter now?

A panel of identity practitioners calls the Model Context Protocol (MCP), with its OAuth layer, the de facto protocol for AI agents' tool calls [BI-FIND-2026-0069](https://insights.bostonidentity.com/ref/BI-FIND-2026-0069). As of its 28 July 2026 revision, an MCP client is an OAuth 2.1 client and an MCP server is an OAuth 2.1 resource server [BI-FIND-2026-0189](https://insights.bostonidentity.com/ref/BI-FIND-2026-0189). Authorization in MCP is optional, and nothing in the token says which person the agent is acting for [BI-FIND-2026-0189](https://insights.bostonidentity.com/ref/BI-FIND-2026-0189). Two identity product builders told one show that vendors under pressure wrap the API they already have, so the fine-grained scopes an agent needs were never there to expose [BI-FIND-2026-0276](https://insights.bostonidentity.com/ref/BI-FIND-2026-0276).

Stolen tokens are already in the incident record. In the public account of the April 2026 Vercel breach, an infostealer at the AI vendor Context.ai exposed OAuth tokens it stored, and those opened a Vercel employee's Google account [BI-FIND-2026-0263](https://insights.bostonidentity.com/ref/BI-FIND-2026-0263). Consent can be phished as well. Datadog Security Labs demonstrated a Copilot Studio agent that sends a victim to the consent screen of an attacker's app and forwards the token once the victim approves; Microsoft treats it as social engineering rather than a vulnerability [BI-FIND-2026-0482](https://insights.bostonidentity.com/ref/BI-FIND-2026-0482).

OAuth is also moving into written requirements for agents. OWASP's AI security verification standard, AISVS 1.0, has MCP servers validate the access token on every request under OAuth 2.1 and never pass a client's token on to downstream APIs [BI-FIND-2026-1473](https://insights.bostonidentity.com/ref/BI-FIND-2026-1473). OAuth 2.1 is still a draft that drops the implicit and password grants [BI-FIND-2012-0005](https://insights.bostonidentity.com/ref/BI-FIND-2012-0005). Dick Hardt, who says he is working on it, says it adds no new features and does not make proof of possession required [BI-FIND-2026-1523](https://insights.bostonidentity.com/ref/BI-FIND-2026-1523).

## Where does it fit?

- Compared with JWT: A JSON Web Token (JWT) is a token format, not a way to get a token. RFC 6749 lets an access token be an identifier or a signed string that carries the authorization itself [S1 §1.4], and a common JWT format for access tokens came later, in RFC 9068 [BI-FIND-2012-0007](https://insights.bostonidentity.com/ref/BI-FIND-2012-0007).
- Compared with Proof of possession: Proof of possession makes the client prove it holds a key when it uses a token. RFC 6749 has no such check; the standards that add it, mutual TLS in RFC 8705 and DPoP in RFC 9449, are separate and optional [BI-FIND-2012-0004](https://insights.bostonidentity.com/ref/BI-FIND-2012-0004).
- Compared with Model Context Protocol: The Model Context Protocol connects AI assistants to tools and reuses OAuth for authorization: its client is an OAuth 2.1 client and its server an OAuth 2.1 resource server [BI-FIND-2026-0189](https://insights.bostonidentity.com/ref/BI-FIND-2026-0189).
- Compared with OAuth Identity and Authorization Chaining Across Domains: RFC 6749's delegation is one hop, from a resource owner to one client [BI-FIND-2012-0008](https://insights.bostonidentity.com/ref/BI-FIND-2012-0008). Carrying authorization further relies on extension grants such as token exchange (RFC 8693), which plug into the hook RFC 6749 left for new grant types [BI-FIND-2012-0008](https://insights.bostonidentity.com/ref/BI-FIND-2012-0008).

## What goes wrong?

- A bearer token, the kind RFC 6750 defines, works for whoever holds it [BI-FIND-2012-0001](https://insights.bostonidentity.com/ref/BI-FIND-2012-0001). RFC 6749 gives a resource server no way to check that the client presenting a token is the client it was issued to [BI-FIND-2012-0004](https://insights.bostonidentity.com/ref/BI-FIND-2012-0004). RFC 9700, published in 2025, says servers should bind access tokens to a key the client holds, a should rather than a must [BI-FIND-2012-0004](https://insights.bostonidentity.com/ref/BI-FIND-2012-0004).
- Two of the four grants carry warnings. The implicit grant hands the access token over in the browser redirect without authenticating the client [S1 §1.3.2]. The RFC says the password grant keeps the password anti-pattern OAuth set out to avoid, and RFC 9700 now says it must not be used [BI-FIND-2012-0005](https://insights.bostonidentity.com/ref/BI-FIND-2012-0005).
- Many protections in the 2012 text are only recommendations. Transport Layer Security (TLS) on the redirection endpoint is a should, because requiring it then was a significant hurdle for many developers [S1 §3.1.2.1]. RFC 6749 requires simple string comparison of redirect URIs only when the full URI was registered, and mentions refresh token rotation only as an example [BI-FIND-2012-0006](https://insights.bostonidentity.com/ref/BI-FIND-2012-0006). RFC 9700 later required exact redirect matching, and rotated or key-bound refresh tokens for public clients [BI-FIND-2012-0006](https://insights.bostonidentity.com/ref/BI-FIND-2012-0006).
- A public client, such as an app running in a browser or on a phone, cannot keep a secret [S1 §2.1]. For such a client, the server checks the code only against a client identifier that the RFC says is not a secret [BI-FIND-2012-0006](https://insights.bostonidentity.com/ref/BI-FIND-2012-0006).
- People often approve what they cannot judge. RFC 6819 makes the end user part of the security model while conceding that users often cannot judge what they approve, and it offers no fix for malicious apps a user installs [BI-FIND-2013-0008](https://insights.bostonidentity.com/ref/BI-FIND-2013-0008).
- Interoperability is not built in. The RFC warns it is likely to produce a wide range of non-interoperable implementations, and it leaves client registration and endpoint discovery partially or fully undefined [S1 §1.8].

## What should a security leader ask?

1. Which third-party apps and AI agents hold OAuth tokens to our core systems today, and with what scopes?
2. Are our access tokens bearer tokens, or bound to a key the client must prove it holds?
3. Which scopes can an ordinary employee grant to an app without an administrator's approval?
4. Do any of our applications still use the implicit or password grant?
5. How long do our refresh tokens last, and can we revoke one app's tokens without touching the others?

## Sources

- [S1] [The OAuth 2.0 Authorization Framework (RFC 6749)](https://www.rfc-editor.org/rfc/rfc6749.txt) (IETF, Standards Track RFC, October 2012)

## Evidence

- [BI-FIND-2012-0001](https://insights.bostonidentity.com/ref/BI-FIND-2012-0001): RFC 6750 defines a bearer token as one that any party holding it can use, with no need to prove possession of a cryptographic key, and the RFC specifies no way to tie a token to its holder. We read its security advice as one long effort to stop leaks and limit the damage when they happen. Jones and Hardt call keeping tokens from unintended parties "the primary security consideration". Clients MUST use TLS and validate the resource server's certificate chain, tokens MUST NOT be stored in cookies sent in the clear, and token lifetimes MUST be limited. Token servers SHOULD issue tokens that last one hour or less and carry an audience restriction. RFC 9700, which updates it, adds key binding: authorization and resource servers SHOULD sender-constrain access tokens, binding each to a key the client holds through mutual TLS (RFC 8705) or DPoP (RFC 9449). (from "RFC 6750 tokens work for whoever holds them; RFC 9700 now urges key binding and bars them from query strings")
- [BI-FIND-2026-0484](https://insights.bostonidentity.com/ref/BI-FIND-2026-0484): Asked what they build today when a client wants people and embedded third-party agents in the same system, Zitel said it was "sticking to the basics": OpenID Connect for sign-in and OAuth for authorization, with delegated scopes, de-scoping and on-behalf-of flows ("that's what we're actually building"). Bonifield described what he sees inside clients: MCP gateways multiplying, from data platforms such as Databricks and proxies such as LiteLLM, and clients asking whether the real single sign-on is "a gateway of gateways"; he would start with visibility across the gateways and identity providers and consolidate later. No client is named. (from "An integrator builds agent access on OpenID Connect and OAuth; clients want one view across many MCP gateways")
- [BI-FIND-2012-0005](https://insights.bostonidentity.com/ref/BI-FIND-2012-0005): Two of the four grants RFC 6749 defines come with warnings in its own security section. The implicit grant returns the access token in the redirect with no client authentication (§4.2), and §10.16 says any sign-in protocol built on OAuth must not use it without extra checks, such as audience-restricting the token. Of the password grant, §10.7 says it "carries a higher risk than other grant types because it maintains the password anti-pattern this protocol seeks to avoid," and asks servers and clients to minimize its use. Thirteen years on, in the later documents we have read, RFC 9700 says the password grant MUST NOT be used and clients SHOULD NOT use the implicit grant, and the OAuth 2.1 draft leaves both out. (from "RFC 6749 admits a server can't tell who presents a token; the fixes arrived in 2020 and 2023")
- [BI-FIND-2012-0007](https://insights.bostonidentity.com/ref/BI-FIND-2012-0007): RFC 6749 warns that on its own it "is likely to produce a wide range of non-interoperable implementations" (§1.8). It names client registration, authorization server capabilities and endpoint discovery as examples of required pieces it leaves undefined, and it puts how a resource server validates a token (§7) and how the authorization and resource servers talk to each other (§1.1) out of scope. Among the RFCs we have read, later standards filled several of these gaps: dynamic client registration in RFC 7591 (2015), server metadata for capabilities and discovery in RFC 8414 (2018), and a common JWT format for access tokens in RFC 9068 (2021). (from "RFC 6749 admits a server can't tell who presents a token; the fixes arrived in 2020 and 2023")
- [BI-FIND-2026-0069](https://insights.bostonidentity.com/ref/BI-FIND-2026-0069): MCP with its community-added OAuth layer is the de facto tool-call protocol, and the panel is unanimous that it is not a policy system: it says how a client gets a token, not who is allowed to do what. The access decision is being pulled up into identity providers and agent gateways that sit above the protocol, which is exactly where Okta, Microsoft Entra and Ping Identity already sit for humans. (from "The rules cover one agent acting for one person; they break when that agent hands work to other agents")
- [BI-FIND-2026-0189](https://insights.bostonidentity.com/ref/BI-FIND-2026-0189): As of the 28 July 2026 revision, an MCP client is an OAuth 2.1 client and an MCP server is an OAuth 2.1 resource server, and that is the whole identity model. Authorization is optional; HTTP transports SHOULD use it, and 'implementations using an STDIO transport SHOULD NOT follow this specification, and instead retrieve credentials from the environment'. The authorization server is out of scope. There is no actor claim, no on-behalf-of flow, no token exchange and no transaction-token profile: nothing in the token says whom the agent is acting for, which is what the Identiverse summit panel meant by 'not a policy system' and what WorkOS meant by the gateway being undefined. Anything of that kind is routed to the new extension model, optional and additive, in a separate repository. This page is the baseline against which the summit's forecast, that the specification gains a delegated-identity mechanism within a year, will be graded. (from "MCP's authorization spec makes the agent a plain OAuth client and leaves delegation to an extension")
- [BI-FIND-2026-0276](https://insights.bostonidentity.com/ref/BI-FIND-2026-0276): The Model Context Protocol specifies OAuth for authorization and leaves it optional, and the people who build identity products say most servers take the option. On this show the head of SecureAuth and the engineer who open-sourced Cedar agree on the mechanism: a software vendor under pressure to ship an MCP server wraps the API it already has, so the fine-grained scopes an agent needs were never there to expose. Their phrase is that it is a big, fat, raw API play. They name no server and give no count. The specification text we read confirms the opening: authorization is optional, and a client that does use it is a plain OAuth client with no delegation. (from "MCP made authorization optional, and two identity veterans say most servers took the option")
- [BI-FIND-2026-0263](https://insights.bostonidentity.com/ref/BI-FIND-2026-0263): The April 2026 Vercel breach is turning into the vendor reference case for agent tokens, and it is growing in the telling. Kahn describes an unnamed vibe-coding company that lost its entire source code through a long-lived, over-privileged token belonging to a third-party agent. The public account is that an infostealer at the AI vendor Context.ai exposed OAuth tokens it stored, which opened a Vercel employee's Google account, and that Vercel lost some source code, environment variables and a limited set of customer credentials. The mechanism he names is accurate. The scale is not what was reported. (from "Okta's answer to agent sprawl: put the identity provider between every agent and every app")
- [BI-FIND-2026-0482](https://insights.bostonidentity.com/ref/BI-FIND-2026-0482): CoPhish, published by Datadog Security Labs on 20 October 2025, turns a Copilot Studio agent into a phishing page that lives on copilotstudio.microsoft.com. The attacker builds an agent, points its sign-in step at the consent screen of an app the attacker registered, and adds an action that forwards the token from Microsoft's servers once the victim approves, so nothing suspicious crosses the victim's browser and MFA never comes into it. Datadog's demo app asked for Mail.ReadWrite, Mail.Send and Notes.ReadWrite, which an ordinary user could still grant under Microsoft's default consent policy; an application administrator could be talked into granting anything. Microsoft treats it as social engineering rather than a vulnerability. (from "A phishing agent hosted on Microsoft's own domain, and why giving agents identities does not stop it")
- [BI-FIND-2026-1473](https://insights.bostonidentity.com/ref/BI-FIND-2026-1473): In the published AISVS 1.0, the requirements an identity team would own sit in three chapters. The agentic chapter (C9) asks for a unique cryptographic identity per agent instance, a scope-limited token that carries the user's authorization to every downstream call, and access decisions made by application logic or a policy engine, "never by the AI model itself". The MCP chapter (C10) has servers validate the access token on each request under OAuth 2.1, check both the tool and its argument values on every invocation, and never pass a client's token through to downstream APIs. The Access Control and Identity chapter (C5) holds 11 requirements, among them step-up authentication for risky AI operations and the end user's permissions enforced inside retrieval pipelines. Levels decide who has to meet these: the agent identity, delegation and per-invocation rules above are Level 2, C10's basic token checks are Level 1, and C5's step-up and federated agent token rules are Level 3, so an assessment at Level 1 covers little of this. On stage, Jim Manico gave the reason MCP has its own chapter: "models don't make security decisions." (from "From Level 2 up, OWASP's AI security standard asks for an identity per agent and a check on every MCP tool call")
- [BI-FIND-2026-1523](https://insights.bostonidentity.com/ref/BI-FIND-2026-1523): Hardt, who says he is working on OAuth 2.1, says it adds no new features and does not make proof of possession a required piece either. On the web, he says, TLS lets a client know which server it reached, but "the server has no idea who's calling it". He traces the wider problem to OAuth being "a framework as opposed to a protocol", with no conformance on what a client must do to work with a given server. In AAuth every HTTP call is signed, so the server knows the caller, and he says the same signing could give webhooks a standard alternative to the shared secrets they rely on today. (from "AAuth carries the signals an agent's approval needs, Hardt says; the brain that decides does not exist yet")
- [BI-FIND-2012-0004](https://insights.bostonidentity.com/ref/BI-FIND-2012-0004): RFC 6749 gives a resource server no way to tell whether the client presenting an access token is the client it was issued to. Section 10.3 says so in as many words: the specification "does not provide any methods for the resource server to ensure that an access token presented to it by a given client was issued to that client by the authorization server." It leaves token types to companion documents (§7.1), and the companion published with it, RFC 6750, defines the bearer token that works for whoever holds it. We read the later binding standards as answers to this gap, each published as a separate, optional document: certificate-bound tokens over mutual TLS in RFC 8705 (2020), DPoP in RFC 9449 (2023), and in RFC 9700 (2025) a SHOULD, not a MUST, that servers sender-constrain access tokens. (from "RFC 6749 admits a server can't tell who presents a token; the fixes arrived in 2020 and 2023")
- [BI-FIND-2012-0008](https://insights.bostonidentity.com/ref/BI-FIND-2012-0008): RFC 6749's delegation is one hop. A resource owner approves a third-party client, and the authorization server issues that client an access token (§1); or a confidential client gets a token with only its own credentials, which the RFC says is typically the client acting on its own behalf (§1.3.4, §4.4). The RFC's text uses the word "agent" only inside "user-agent", which it describes as typically a web browser. It does define the hook that later delegation work plugs into: an extension grant is any absolute URI sent as grant_type (§4.5). The JWT bearer grant (RFC 7523) and token exchange (RFC 8693) are grants of that kind. Our briefing on the 2026 identity-assertion grant draft found that it profiles both, and its appendix walks through an AI agent as the OAuth client. (from "RFC 6749 admits a server can't tell who presents a token; the fixes arrived in 2020 and 2023")
- [BI-FIND-2012-0006](https://insights.bostonidentity.com/ref/BI-FIND-2012-0006): Several protections that later RFCs tightened sit in RFC 6749 as conditions or examples. Only public clients and confidential clients using the implicit grant must register a redirect URI; for others it is a SHOULD (§3.1.2.2). The server should ask for the complete URI, and if it cannot, it should at least require scheme, authority and path. Simple string comparison is required when the full URI was registered; otherwise the server matches against the registered parts (§3.1.2.3). For a public client, the code is checked against the client_id in the request (§4.1.3), which the RFC says is not a secret (§2.2). Even TLS on the redirection endpoint is a SHOULD, because in 2012 requiring it was "a significant hurdle for many client developers" (§3.1.2.1). Refresh token rotation appears in §10.4 only as an example of how a server could detect abuse when it cannot authenticate the client. Among the later RFCs we have read, PKCE (RFC 7636, 2015) has the client prove at the token endpoint that it holds a one-time key it created for that authorization request, and RFC 9700 (2025) requires exact redirect matching and sender-constrained or rotated refresh tokens for public clients. (from "RFC 6749 admits a server can't tell who presents a token; the fixes arrived in 2020 and 2023")
- [BI-FIND-2013-0008](https://insights.bostonidentity.com/ref/BI-FIND-2013-0008): RFC 6819 does not rank its threats. It says it cannot assess the probability or risk of a threat, because both depend on the application and deployment, and offers itself as a foundation for deployment-specific threat models. It also makes the end user part of the security model, while conceding that users often cannot judge what they approve. For malicious user-installed apps that steal credentials, over-request scope or approve on the user's behalf, section 5.5 offers no fix: "No solution is offered here, because none is known." (from "RFC 6819 warned in 2013 that servers receiving a token may reuse it, and offered audience binding and client-bound tokens")

---

© 2026 Boston Identity LLC. All rights reserved. Published by Boston Identity LLC.
Quote up to 100 words with the credit "Boston Identity, IAM Insights" and a link to https://insights.bostonidentity.com/glossary/oauth. Claim numbers may be cited freely. Training, fine-tuning and text-and-data-mining use of this page is permitted.
Terms: https://insights.bostonidentity.com/terms · How to cite: https://insights.bostonidentity.com/cite
