# What is CIBA?

> CIBA is an OpenID sign-in flow that lets an app ask a provider to authenticate a known user on a separate device, usually their phone, and then collect the tokens once the user approves, without any browser redirect.

Explainer by Boston Identity, 2026-10-20. 6 min read. https://insights.bostonidentity.com/glossary/ciba

- What it is: protocol
- Defined by: OpenID Foundation, Final (2021)
- Who it's for: Teams adding human approval to agents

## What is CIBA?

Client-Initiated Backchannel Authentication (CIBA) is an OpenID Connect flow in which the app starts the sign-in, not the user's browser. OpenID Connect is an identity layer that lets an app, called the Relying Party, confirm who its user is through an OpenID Provider. In its usual flows the user's own browser is redirected to the provider and back. CIBA drops the redirect: the app, called the client, talks to the OpenID Provider directly over a back channel [S1 §1].

The spec splits the user's world into two devices. The consumption device is where the user meets the service, and the user need not control it: a bank teller's screen or a petrol pump are the spec's examples. The authentication device, often a smartphone, is where the user signs in and approves [S1 §2]. The client must already hold an identifier for the user, such as a phone number, email address or earlier ID token [S1 §7.1]. The spec's own use cases are a call centre checking a caller, a teller checking a customer in a branch, and a phone approving a payment at a shop terminal [S1 §6].

Think of it like this: It works like a shop assistant who rings your bank while you stand at the till, and the bank sends a prompt to your phone; the assistant waits for the bank's answer, not yours. The comparison stops at what the prompt says: CIBA's optional shared message is meant as a short code that matches the two screens, not a description of what you are approving.

Not the same as: CIBA is not a push-notification product or a second factor in itself. The spec leaves how the provider reaches the user and authenticates them to the provider; it only defines how the client asks and how the result comes back.

## How does CIBA work?

Three parties take part: the client (the app that wants the user signed in), the OpenID Provider, and the user's authentication device. Before any of this, the client registers one way of receiving results: poll, ping or push [S1 §4].

[Figure: The client never redirects the user: it asks the OpenID Provider over a back channel, the provider asks the user on their phone, and the client collects the tokens with the auth_req_id it was given.]

1. **The client asks the provider to authenticate a user.** The client sends an HTTP POST straight to the provider's backchannel authentication endpoint and authenticates itself as it would at the token endpoint. The request must carry the openid scope and exactly one hint naming the user: a login_hint_token, an earlier ID token, or a plain login_hint. A ping or push client must also include a client_notification_token, a random bearer token the provider will later use to prove its callbacks are genuine. The request may add a binding_message, a short code shown on both devices, and a user_code, a secret the user supplies [S1 §7.1].
2. **The provider checks the request and replies at once.** The provider authenticates the client, validates the parameters, and must reject a request with more than one hint. It then checks that the hint points to a real user. If all is well it returns HTTP 200 with an auth_req_id, a hard-to-guess identifier for this request, and how long that identifier lives. For poll and ping clients it may also return an interval, the minimum wait between token requests; without one, clients wait five seconds [S1 §7.2, §7.3].
3. **The provider reaches the user on their own device.** The provider identifies the user and picks a channel to authenticate them, usually by opening an interactive session on the authentication device. Once the user is authenticated, the provider must obtain their decision to authorize the request, as in ordinary OpenID Connect. There is no dialogue between the provider and the user through the consumption device, though a member of the client's staff may read the binding_message out to the user [S1 §8].
4. **The client waits for the result.** A poll client calls the token endpoint with the CIBA grant type and the auth_req_id, no more often than the interval allows. Until the user acts it gets authorization_pending, and slow_down means it must wait at least five seconds longer. A ping client instead receives a callback carrying only the auth_req_id, then calls the token endpoint [S1 §10.1, §10.2, §11].
5. **The provider hands over the tokens.** When the user has authenticated and approved, the provider returns an ID token, an access token and optionally a refresh token, and the auth_req_id stops working. If the user refused, the client gets access_denied. A push client receives the tokens directly at its notification endpoint, and the ID token must carry the auth_req_id and a hash of the access token, which tie the delivery together. The client must check the client_notification_token on every callback [S1 §10.1.1, §10.3.1, §11].

## Why does CIBA matter now?

The spec's own examples are call centres, bank branches and shop terminals, and several agent designs now use CIBA when an AI agent needs a person's approval. In a Ping Identity demo of banking through ChatGPT, Ping's MCP gateway asked the person to approve the agent's action through Ping's Recognize or through CIBA on their own device [BI-FIND-2026-0011](https://insights.bostonidentity.com/ref/BI-FIND-2026-0011). Highflame's open-source ZeroID issues a token naming the approving person as subject and the agent in a separate claim after a CIBA approval [BI-FIND-2026-0139](https://insights.bostonidentity.com/ref/BI-FIND-2026-0139). An IETF draft on agent authorization points to CIBA for human approval [BI-FIND-2026-0181](https://insights.bostonidentity.com/ref/BI-FIND-2026-0181).

The fit is loose, and the people writing standards have noticed. That same draft notes CIBA only covers requests the client starts and maps poorly to approval needed mid-task [BI-FIND-2026-0181](https://insights.bostonidentity.com/ref/BI-FIND-2026-0181). An OAuth working-group draft lets the server decide at the token endpoint to wait for a person, because agents mostly use grants CIBA cannot hold [BI-FIND-2026-0629](https://insights.bostonidentity.com/ref/BI-FIND-2026-0629). OAuth's draft cross-device guidance ranks CIBA second, behind FIDO and ahead of the device-code flow [BI-FIND-2026-1079](https://insights.bostonidentity.com/ref/BI-FIND-2026-1079).

## Where does it fit?

- Compared with OpenID Connect: CIBA keeps OpenID Connect's meaning and tokens but adds a new endpoint and an asynchronous way to deliver the result, with no redirect through the user's browser [S1 §1].
- Compared with OAuth 2.0 Device Authorization Grant: Both make a client wait while a person approves elsewhere, and CIBA reuses the device flow's polling errors [S1 §11]. OAuth's draft cross-device guidance rates CIBA less exposed than the device grant, though still open to anyone who knows the user's identifier [BI-FIND-2026-1081](https://insights.bostonidentity.com/ref/BI-FIND-2026-1081).
- Compared with OAuth Deferred Token Response: CIBA is started by the client, which must name the user up front. The deferred-token draft lets the authorization server decide, at the token endpoint, to wait and choose whom to ask [BI-FIND-2026-0629](https://insights.bostonidentity.com/ref/BI-FIND-2026-0629).

## What goes wrong?

- A registered client, or a person working at one, who knows a user's identifier can start an approval prompt on that user's phone. The user_code guard against unsolicited prompts is optional for providers, and the spec says only that the login_hint_token should be signed [S1 §7.1.2, §14] [BI-FIND-2021-0002](https://insights.bostonidentity.com/ref/BI-FIND-2021-0002). OAuth's draft cross-device guidance gives hypothetical cases of attackers sending requests to a bank's user list until a few approve, and flooding an administrator at night until one is approved by mistake [BI-FIND-2026-1081](https://insights.bostonidentity.com/ref/BI-FIND-2026-1081).
- The person may approve without seeing the action. The binding_message is designed as a short matching code, the spec defines no field for an action or amount, and no token carries the message back [S1 §7.1] [BI-FIND-2021-0001](https://insights.bostonidentity.com/ref/BI-FIND-2021-0001). We read that as meaning an agent design that uses CIBA to approve a specific act needs a profile on top.
- Push mode skips the token endpoint, where clients normally authenticate before tokens are released. The spec asks implementers to secure the notification endpoint and suggests sender-constrained tokens against interception [S1 §14]. The provider should also check that the registered endpoint really belongs to the client, but the spec does not say how [S1 §14].
- An approval step only helps if something waits for it. In one Auth0 guide a prompt instruction decides when to call CIBA, and the agent told the user a password was reset when no reset happened [BI-FIND-2026-1261](https://insights.bostonidentity.com/ref/BI-FIND-2026-1261).

## What should a security leader ask?

1. When one of our agents asks a person to approve an action through CIBA, what does the phone show besides a short code?
2. Can anyone who knows a customer's phone number or email trigger an approval prompt, and have we turned on user codes or rate limits?
3. Which of our CIBA clients use push mode, and who controls the notification endpoints they registered?
4. Which agent actions actually wait for the token from an approval, rather than for a model's decision to ask?
5. Did we do a written risk assessment before enabling cross-device approval, as OAuth's draft guidance says implementers must?

## Sources

- [S1] [OpenID Connect Client-Initiated Backchannel Authentication Flow - Core 1.0](https://openid.net/specs/openid-client-initiated-backchannel-authentication-core-1_0.html) (OpenID Foundation, Final, 1 September 2021)

## Evidence

- [BI-FIND-2026-0011](https://insights.bostonidentity.com/ref/BI-FIND-2026-0011): Ping demoed headless banking through ChatGPT. A Ping MCP gateway handled OAuth discovery and the user's consent for the agent to act. It also asked a person to approve the action, through Recognize or through Client-Initiated Backchannel Authentication (CIBA), which sends the approval request to the user's own device. LPL Financial's AI assistant is live for about 30,000 advisors who manage $2.3 trillion in assets. (from "Ping moved agent identity from roadmap to shipping product, and let AI agents run its admin tools too")
- [BI-FIND-2026-0139](https://insights.bostonidentity.com/ref/BI-FIND-2026-0139): Human approval is written into the token rather than kept in a separate approval system: an agent sends a CIBA request with a binding message, the person approves out of band, and the token ZeroID issues names the approving person as its subject and the asking agent in a backchannel_client_id claim. The README pitches this as a record that the agent acted with the person's explicit approval at a stated time, with the binding message; its token example shows only the two identities, and by our reading of CIBA core, the spec returns no binding message in any token. Rich authorization requests carry the scope detail into the approval prompt. The record is only as good as the deployer's gateway: ZeroID has no user login of its own, and the README warns that whoever can reach the approve endpoint "approves as anyone". In our reading it is the consent trail the summit panel and Mayer Brown's lawyer asked for, in a vendor's design rather than anyone's deployment. (from "Highflame's open-source ZeroID narrows an agent's token at every hop, a step we did not find in Uber's write-up")
- [BI-FIND-2026-0181](https://insights.bostonidentity.com/ref/BI-FIND-2026-0181): The draft takes its authorization model from OAuth's existing grants. A user delegates through the authorization code grant and SHOULD sign in with a phishing-resistant method such as a passkey; an agent acting for itself can use client credentials or a JWT authorization grant. Where an LLM, Tool or Agent is built from several microservices, the Agent, LLM or Tools can exchange the access token they received for a transaction token, which is bound to one transaction and passed along the internal call chain. For human approval the draft points to CIBA and says "the agent MUST NOT treat local UI confirmation alone as sufficient authorization", then notes that CIBA only accounts for client initiation and maps poorly to confirmation needed mid-execution. A Tool forwarding the access token it got from the Agent is called an anti-pattern in a note, with no MUST NOT. Our briefing on the MCP authorization specification found that spec states its rule against passing tokens on as a MUST NOT. (from "Six companies' agent-auth draft, since adopted by the WIMSE group, keeps credentials out of the model's reach")
- [BI-FIND-2026-0629](https://insights.bostonidentity.com/ref/BI-FIND-2026-0629): Agents mostly get tokens through client credentials and token exchange, and those are the grants CIBA cannot hold for a person's decision: by the authors' account it is client-initiated and needs an identified user up front. This draft lets the authorization server decide at the token endpoint to wait and pick whom to ask, and its own agent use case has the server contact the human principal when an agent exceeds its pre-approved scopes. Two other drafts from different author groups point at the same missing piece, Rosomakho et al. with polling-based approval of one agent operation and Chen et al. listing the lack of an interactive channel among their six gaps. (from "OAuth adopts a way for any token request, including an agent's, to wait for a person's decision")
- [BI-FIND-2026-1079](https://insights.bostonidentity.com/ref/BI-FIND-2026-1079): The draft's protocol-selection summary ranks three options. FIDO cross-device authentication, the CTAP 2.2 "hybrid transports" flow that lets a phone act as the authenticator for signing in on another device, "provides the best protection against attacks on the unauthenticated channel". It proves the two devices are close with Bluetooth Low Energy advertisements, while an authorization server, the draft notes, "cannot independently measure or enforce proximity on its own." FIDO2/WebAuthn SHOULD be used whenever the devices are capable and the device being signed in to has no suitable FIDO credential of its own. If FIDO is not available, CIBA is the alternative, though it is still open to attackers who know or can guess a user's identifier. The OAuth device authorization grant (RFC 8628) comes last: it is RECOMMENDED only with additional mitigations, and the draft's guidance is to use it only when other protocols are not viable and to avoid it where the protected resources are "sensitive, high value, or business critical." (from "OAuth's cross-device guidance ranks FIDO first, then CIBA backchannel login, and the device-code flow last")
- [BI-FIND-2026-1081](https://insights.bostonidentity.com/ref/BI-FIND-2026-1081): Three of the draft's hypothetical examples show push approvals, the backchannel pattern CIBA uses, being abused where no code changes hands. An attacker with a list of a bank's user identifiers triggers a transaction request for each, and "a small percentage grants authorization". A fake helpdesk caller has a user decline a run of fraudulent requests, then approve one that moves all the funds to a new account. And an attacker floods a network administrator with requests at night until one is approved by accident. The draft ties the first of these to its rating of CIBA: less susceptible than the device grant, but "still vulnerable to attackers who know or can guess the user identifier". Of its mitigations, rate limits slow such attacks but do not prevent targeted ones, and authenticate-then-initiate lets the server prevent an attacker from starting the request at all, though it needs a device able to run phishing-resistant sign-in, which the draft says may remove the need for a cross-device flow. (from "OAuth's cross-device guidance ranks FIDO first, then CIBA backchannel login, and the device-code flow last")
- [BI-FIND-2021-0002](https://insights.bostonidentity.com/ref/BI-FIND-2021-0002): CIBA starts on the client's side, so the client must already know whom to ask. Each request carries exactly one user hint (a login_hint_token, an earlier ID token or a plain login_hint), and the provider MUST reject a request with more than one. A client, or a person at a client, who knows a user's identifier can therefore start a prompt to that user, and the spec's guards against that are soft: implementing user_code, a secret the user supplies so others cannot start requests to their device, is optional for providers, and the login_hint_token SHOULD be signed. The spec's own examples are a call-centre agent, a bank teller and a phone approving a payment at a point-of-sale terminal. Its "agent of the Client" who may pass the binding message on is relying-party staff, as in the bank-teller example; the text never mentions software agents. (from "CIBA's human-approval step carries a short matching code, not a description of the action")
- [BI-FIND-2021-0001](https://insights.bostonidentity.com/ref/BI-FIND-2021-0001): In CIBA, the one request parameter meant to be shown to the person is binding_message, and the spec designs it as a short code for matching two screens. It says the value SHOULD be relatively short, in a limited set of plain-text characters, and should let the user tell that the two devices belong to one transaction, "such as a random value of reasonable entropy (e.g. a transactional approval code)." Past that, the provider MUST obtain the person's authorization decision as OpenID Connect Core describes it. The spec defines no parameter for an action or an amount, though it lets extensions and profiles add parameters, and none of the tokens it defines carries the binding message back to the client. Several agent designs we have read cite CIBA for approving a specific action; we read that as needing a profile on top of the core spec. (from "CIBA's human-approval step carries a short matching code, not a description of the action")
- [BI-FIND-2026-1261](https://insights.bostonidentity.com/ref/BI-FIND-2026-1261): Patil warns that "Relying solely on prompt engineering is insufficient for ensuring LLM safety", yet in his build a prompt decides when a person is asked. The agent's system prompt tells the model that for an elevated operation such as a password reset "you MUST call the `invokeCiba` tool", which sends an approval request to the user's Guardian app. The agent code in the post has no reset tool behind the approval (the flow "does not actually reset your password"), so the walkthrough shows an approval arriving on a phone, not an action that waits for one. In the test screenshot the agent still tells the user the new password has been set. (from "Auth0's AgentCore guide says prompts alone are not enough, then lets the prompt decide when to ask a person")

---

© 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/ciba. 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
