GHSA-7qjx-gp9h-65qj

Summary

server/handlers.go::handleTokenExchange (lines 1804-1893) does not call isConnectorAllowed(client.AllowedConnectors, connID) before issuing tokens, while sibling handlers do. This is a per-client connector ACL gap on the token-exchange endpoint; the redirect-flow paths enforce the same field correctly.

Affected code path

handleTokenExchange reads connector_id from the request body at server/handlers.go:1822. Validators called between read and token issuance:

• s.getConnector(ctx, connID) at line 1836 - confirms connector exists
• GrantTypeAllowed(conn.GrantTypes, grantTypeTokenExchange) at line 1842 - confirms connector permits this grant
• (missing) isConnectorAllowed(client.AllowedConnectors, connID) - never called

Tokens are issued at lines 1887 / 1889, bound to client.ID carrying claims derived from connID.

Sibling handlers DO enforce the check:

• server/handlers.go::handleConnectorLogin:377 - calls isConnectorAllowed, returns HTTP 403 "Connector not allowed for this client." (line 380).
• server/oauth2.go::parseAuthorizationRequest:535 - same enforcement for the authorization-code flow.

The doc-string at storage/storage.go:192-194 reads:

AllowedConnectors is a list of connector IDs that the client is allowed to use for authentication. If empty, all connectors are allowed.

The phrasing is unconditional - a permission ACL, not a UX filter.

Impact (concrete scenario)

• Connector corp-okta - high-trust, gates production access
• Connector dev-google - low-trust, internal Gmail
• Client dev-app configured with allowedConnectors: ["dev-google"] (admin intent: dev-app only sees dev-google identities)
• dev-apps client secret leaks (CI artifact, env file, breached service-account secret store)

Without the bug, the leaked secret would only allow the attacker to mint tokens via dev-google - blast radius bounded by what any dev-google user can already do.

With the bug, an attacker holding their own legitimate corp-okta ID token sends:

POST /token
Content-Type: application/x-www-form-urlencoded

grant_type=urn:ietf:params:oauth:grant-type:token-exchange
&client_id=dev-app
&client_secret=<leaked>
&connector_id=corp-okta
&subject_token=<attackers own corp-okta id token>
&subject_token_type=urn:ietf:params:oauth:token-type:id_token
&scope=openid+groups

Dex returns an ID token signed by Dex, aud=dev-app, carrying the attackers corp-okta groups. Downstream services trusting tokens issued for dev-app see the attacker as a corp-okta user - a combination the admins policy explicitly forbade.

Severity (self-assessed)

CVSS 3.1 vector: AV:N/AC:L/PR:H/UI:N/S:C/C:H/I:H/A:N -> 8.7 HIGH.

The PR:H precondition is a real reduction (requires leaked confidential client_secret PLUS attacker holding a subject_token from a forbidden connector that has token-exchange enabled). Defer to your scoring - HIGH and MEDIUM are both defensible.

Affected version

master only - not yet in any released tag. Latest release v2.45.1 (2026-03-03) predates PR #4610 (commit f80a89d, 2026-03-11) which introduced AllowedConnectors. Production deployments on stable releases are NOT affected; deployments pulling from master / nightly images are. A fix can be merged ahead of the next release without an embargo for past versions.

Precedent / lineage

• PR #4610 (commit f80a89d, 2026-03-11) - added the AllowedConnectors field, isConnectorAllowed, filterConnectors, and the redirect-flow check sites (handleConnectorLogin:377, parseAuthorizationRequest:535). Did not modify handleTokenExchange.
• PR #4619 (commit 7777773, 2026-03-11, same author, one day earlier) - added GrantTypeAllowed(conn.GrantTypes, grantTypeTokenExchange) to handleTokenExchange. Added a connector-side grant-type gate but did not add the symmetric client-side connector ACL.

Suggested fix

Insert isConnectorAllowed(client.AllowedConnectors, connID)…

Summary

server/handlers.go::handleTokenExchange (lines 1804-1893) does not call isConnectorAllowed(client.AllowedConnectors, connID) before issuing tokens, while sibling handlers do. This is a per-client connector ACL gap on the token-exchange endpoint; the redirect-flow paths enforce the same field correctly.

Affected code path

handleTokenExchange reads connector_id from the request body at server/handlers.go:1822. Validators called between read and token issuance:

• s.getConnector(ctx, connID) at line 1836 - confirms connector exists
• GrantTypeAllowed(conn.GrantTypes, grantTypeTokenExchange) at line 1842 - confirms connector permits this grant
• (missing) isConnectorAllowed(client.AllowedConnectors, connID) - never called

Tokens are issued at lines 1887 / 1889, bound to client.ID carrying claims derived from connID.

Sibling handlers DO enforce the check:

• server/handlers.go::handleConnectorLogin:377 - calls isConnectorAllowed, returns HTTP 403 "Connector not allowed for this client." (line 380).
• server/oauth2.go::parseAuthorizationRequest:535 - same enforcement for the authorization-code flow.

The doc-string at storage/storage.go:192-194 reads:

AllowedConnectors is a list of connector IDs that the client is allowed to use for authentication. If empty, all connectors are allowed.

The phrasing is unconditional - a permission ACL, not a UX filter.

Impact (concrete scenario)

• Connector corp-okta - high-trust, gates production access
• Connector dev-google - low-trust, internal Gmail
• Client dev-app configured with allowedConnectors: ["dev-google"] (admin intent: dev-app only sees dev-google identities)
• dev-apps client secret leaks (CI artifact, env file, breached service-account secret store)

Without the bug, the leaked secret would only allow the attacker to mint tokens via dev-google - blast radius bounded by what any dev-google user can already do.

With the bug, an attacker holding their own legitimate corp-okta ID token sends:

POST /token
Content-Type: application/x-www-form-urlencoded

grant_type=urn:ietf:params:oauth:grant-type:token-exchange
&client_id=dev-app
&client_secret=<leaked>
&connector_id=corp-okta
&subject_token=<attackers own corp-okta id token>
&subject_token_type=urn:ietf:params:oauth:token-type:id_token
&scope=openid+groups

Dex returns an ID token signed by Dex, aud=dev-app, carrying the attackers corp-okta groups. Downstream services trusting tokens issued for dev-app see the attacker as a corp-okta user - a combination the admins policy explicitly forbade.

Severity (self-assessed)

CVSS 3.1 vector: AV:N/AC:L/PR:H/UI:N/S:C/C:H/I:H/A:N -> 8.7 HIGH.

The PR:H precondition is a real reduction (requires leaked confidential client_secret PLUS attacker holding a subject_token from a forbidden connector that has token-exchange enabled). Defer to your scoring - HIGH and MEDIUM are both defensible.

Affected version

master only - not yet in any released tag. Latest release v2.45.1 (2026-03-03) predates PR #4610 (commit f80a89d, 2026-03-11) which introduced AllowedConnectors. Production deployments on stable releases are NOT affected; deployments pulling from master / nightly images are. A fix can be merged ahead of the next release without an embargo for past versions.

Precedent / lineage

• PR #4610 (commit f80a89d, 2026-03-11) - added the AllowedConnectors field, isConnectorAllowed, filterConnectors, and the redirect-flow check sites (handleConnectorLogin:377, parseAuthorizationRequest:535). Did not modify handleTokenExchange.
• PR #4619 (commit 7777773, 2026-03-11, same author, one day earlier) - added GrantTypeAllowed(conn.GrantTypes, grantTypeTokenExchange) to handleTokenExchange. Added a connector-side grant-type gate but did not add the symmetric client-side connector ACL.

Suggested fix

Insert isConnectorAllowed(client.AllowedConnectors, connID)…

Dex: Token-exchange endpoint is missing AllowedConnectors enforcement in github.com/dexidp/dex