OAuth 2.0 Proof-of-Possession (PoP) Security Architecture

The OAuth 2.0 protocol family (, , and ) offer a single token type known as the "bearer" token to access protected resources. RFC 6750 specifies the bearer token mechanism and defines it as follows: "A security token with the property that any party in possession of the token (a "bearer") can use the token in any way that any other party in possession of it can. Using a bearer token does not require a bearer to prove possession of cryptographic key material." The bearer token meets the security needs of a number of use cases the OAuth 2.0 protocol had originally been designed for. There are, however, other scenarios that require stronger security properties and ask for active participation of the OAuth client in form of cryptographic computations when presenting an access token to a resource server. This document outlines additional use cases requiring stronger security protection in , identifies threats in , proposes different ways to mitigate those threats in , outlines an architecture for a solution that builds on top of the existing OAuth 2.0 framework in , and concludes with a requirements list in .

The key words 'MUST', 'MUST NOT', 'REQUIRED', 'SHALL', 'SHALL NOT', 'SHOULD', 'SHOULD NOT', 'RECOMMENDED', 'MAY', and 'OPTIONAL' in this specification are to be interpreted as described in , with the important qualification that, unless otherwise stated, these terms apply to the design of the protocol, not its implementation or application.

The main use case that motivates improvement upon "bearer" token security is the desire of resource servers to obtain additional assurance that the client is indeed authorized to present an access token. The expectation is that the use of additional credentials (symmetric or asymmetric keying material) will encourage developers to take additional precautions when transferring and storing access token in combination with these credentials. Additional use cases listed below provide further requirements for the solution development. Note that a single solution does not necessarily need to offer support for all use cases.

In a scenario where a resource server receives a valid access token, the resource server then re-uses it with other resource server. The reason for re-use may be malicious or may well be legitimate. In a legitimate case, the intent is to support chaining of computations whereby a resource server needs to consult other third party resource servers to complete a requested operation. In both cases it may be assumed that the scope and audience of the access token is sufficiently defined that to allow such a re-use. For example, imagine a case where a company operates email services as well as picture sharing services and that company had decided to issue access tokens with a scope and audience that allows access to both services. With this use case the desire is to prevent such access token re-use. This also implies that the legitimate use cases require additional enhancements for request chaining.

In this use case we consider the scenario where an OAuth 2.0 request to a protected resource is secured using TLS or DTLS (see ), but the client and the resource server demand that the underlying TLS/DTLS exchange is bound to additional application layer security to prevent cases where the TLS/DTLS connection is terminated at a TLS/DTLS intermediary, which splits the TLS/DTLS connection into two separate connections. In this use case additional information should be conveyed to the resource server to ensure that no entity entity has tampered with the TLS/DTLS connection.

This use case is for a web client that needs to access a resource that makes data available (such as videos) without offering integrity and confidentiality protection using TLS. Still, the initial resource request using OAuth, which includes the access token, must be protected against various threats (e.g., token replay, token modification). While it is possible to utilize bearer tokens in this scenario with TLS protection when the request to the protected resource is made, as described in , there may be the desire to avoid using TLS between the client and the resource server at all. In such a case the bearer token approach is not possible since it relies on TLS for ensuring integrity and confidentiality protection of the access token exchange since otherwise replay attacks are possible: First, an eavesdropper may steal an access token and present it at a different resource server. Second, an eavesdropper may steal an access token and replay it against the same resource server at a later point in time. In both cases, if the attack is successful, the adversary gets access to the resource owners data or may perform an operation selected by the adversary (e.g., sending a message). Note that the adversary may obtain the access token (if the recommendations in and are not followed) using a number of ways, including eavesdropping the communication on the wireless link. Consequently, the important assumption in this use case is that a resource server does not have TLS support and the security solution should work in such a scenario. Furthermore, it may not be necessary to provide authentication of the resource server towards the client.

In Web deployments resource servers are often placed behind load balancers, which are deployed by the same organization that operates the resource servers. These load balancers may terminate the TLS connection setup and HTTP traffic is transmitted without TLS protection from the load balancer to the resource server. With application layer security in addition to the underlying TLS security it is possible to allow application servers to perform cryptographic verification on an end-to-end basis. The key aspect in this use case is therefore to offer end-to-end security in the presence of load balancers via application layer security. Enterprise networks also deploy proxies that inspect traffic and thereby break TLS.

The following list presents several common threats against protocols utilizing some form of token. This list of threats is based on NIST Special Publication 800-63 . We exclude a discussion of threats related to any form of identity proofing and authentication of the resource owner to the authorization server since these procedures are not part of the OAuth 2.0 protocol specification itself. An attacker may generate a bogus token or modify the token content (such as authentication or attribute statements) of an existing token, causing resource server to grant inappropriate access to the client. For example, an attacker may modify the token to extend the validity period. A client, which MAY be a normal client or MAY be assumed to be constrained (see ), may modify the token to have access to information that they should not be able to view. Tokens may contain personal data, such as real name, age or birthday, payment information, etc. An attacker uses the token generated for consumption by the resource server to obtain access to another resource server. An attacker attempts to use a token that has already been used once with a resource server. The attacker may be an eavesdropper who observes the communication exchange or, worse, one of the communication end points. A client may, for example, leak access tokens because it cannot keep secrets confidential. A client may also reuse access tokens for some other resource servers. Finally, a resource server may use a token it had obtained from a client and use it with another resource server that the client interacts with. A resource server, offering relatively unimportant application services, may attempt to use an access token obtained from a client to access a high-value service, such as a payment service, on behalf of the client using the same access token. Token repudiation refers to a property whereby a resource server is given an assurance that the authorization server cannot deny to have created a token for the client.

RFC 4962 gives useful guidelines for designers of authentication and key management protocols. While RFC 4962 was written with the AAA framework used for network access authentication in mind the offered suggestions are useful for the design of other key management systems as well. The following requirements list applies OAuth 2.0 terminology to the requirements outlined in RFC 4962. These requirements include The key management protocol MUST be cryptographic algorithm independent. Session keys MUST be strong and fresh. Each session deserves an independent session key, i.e., one that is generated specifically for the intended use. In context of OAuth this means that keying material is created in such a way that can only be used by the combination of a client instance, protected resource, and authorization scope. Following the principle of least privilege, parties MUST NOT have access to keying material that is not needed to perform their role. Any protocol that is used to establish session keys MUST specify the scope for session keys, clearly identifying the parties to whom the session key is available. The key management protocol exchanges MUST be replay protected. Replay protection allows a protocol message recipient to discard any message that was recorded during a previous legitimate dialogue and presented as though it belonged to the current dialogue. Each party in the key management protocol MUST be authenticated to the other parties with whom they communicate. Authentication mechanisms MUST maintain the confidentiality of any secret values used in the authentication process. Secrets MUST NOT be sent to another party without confidentiality protection. Client and resource server authorization MUST be performed. These entities MUST demonstrate possession of the appropriate keying material, without disclosing it. Authorization is REQUIRED whenever a client interacts with an authorization server. Authorization checking prevents an elevation of privilege attack. While preserving algorithm independence, confidentiality and integrity of all keying material MUST be maintained. The selection of the "best" cryptographic algorithms SHOULD be securely confirmed. The mechanism SHOULD detect attempted roll-back attacks. Key management proposals require a robust key naming scheme, particularly where key caching is supported. The key name provides a way to refer to a key in a protocol so that it is clear to all parties which key is being referenced. Objects that cannot be named cannot be managed. All keys MUST be uniquely named, and the key name MUST NOT directly or indirectly disclose the keying material. Compromise of a single client MUST NOT compromise keying material held by any other client within the system, including session keys and long-term keys. Likewise, compromise of a single resource server MUST NOT compromise keying material held by any other Resource Server within the system. In the context of a key hierarchy, this means that the compromise of one node in the key hierarchy must not disclose the information necessary to compromise other branches in the key hierarchy. Obviously, the compromise of the root of the key hierarchy will compromise all of the keys; however, a compromise in one branch MUST NOT result in the compromise of other branches. There are many implications of this requirement; however, two implications deserve highlighting. First, the scope of the keying material must be defined and understood by all parties that communicate with a party that holds that keying material. Second, a party that holds keying material in a key hierarchy must not share that keying material with parties that are associated with other branches in the key hierarchy. Keying material MUST be bound to the appropriate context. The context includes the following. The manner in which the keying material is expected to be used. The other parties that are expected to have access to the keying material. The expected lifetime of the keying material. Lifetime of a child key SHOULD NOT be greater than the lifetime of its parent in the key hierarchy. Any party with legitimate access to keying material can determine its context. In addition, the protocol MUST ensure that all parties with legitimate access to keying material have the same context for the keying material. This requires that the parties are properly identified and authenticated, so that all of the parties that have access to the keying material can be determined. The context will include the client and the resource server identities in more than one form. If client authorization is restricted, then the client SHOULD be made aware of the restriction. A client has identity confidentiality when any party other than the resource server and the authorization server cannot sufficiently identify the client within the anonymity set. In comparison to anonymity and pseudonymity, identity confidentiality is concerned with eavesdroppers and intermediaries. A key management protocol SHOULD provide this property. Resource servers SHOULD be prevented from knowing the real or pseudonymous identity of the resource owner, since the authorization server is the only entity involved in verifying the resource owner's identity. Resource servers that collude can be prevented from using information related to the resource owner to track the individual. That is, two different resource servers can be prevented from determining that the same resource owner has authenticated to both of them. Authorization servers MUST bind different keying material to access tokens used for resource servers from different origins (or similar concepts in the app world). For solutions using asymmetric key cryptography the client MAY conceal information about the resource server it wants to interact with. The authorization server MAY reject such an attempt since it may not be able to enforce access control decisions. A solution MUST enable support for channel bindings. The concept of channel binding, as defined in , allows applications to establish that the two end-points of a secure channel at one network layer are the same as at a higher layer by binding authentication at the higher layer to the channel at the lower layer. There are performance concerns with the use of asymmetric cryptography. Although symmetric key cryptography offers better performance asymmetric cryptography offers additional security properties. A solution MUST therefore offer the capability to support both symmetric as well as asymmetric keys. There are threats that relate to the experience of the software developer as well as operational practices. Verifying the servers identity in TLS is discussed at length in . A number of the threats listed in demand protection of the access token content and a standardized solution, for example, in the form of a JSON-based format, is available with the JWT .

A large range of threats can be mitigated by protecting the content of the token, for example using a digital signature or a keyed message digest. Alternatively, the content of the token could be passed by reference rather than by value (requiring a separate message exchange to resolve the reference to the token content). To simplify discussion in the following example we assume that the token itself cannot be modified by the client, either due to cryptographic protection (such as signature or encryption) or use of a reference value with sufficient entropy and associated secure lookup. The token remains opaque to the client. These are characteristics shared with bearer tokens and more information on best practices can be found in and in the security considerations section of . To deal with token redirect it is important for the authorization server to include the identifier of the intended recipient - the resource server. A resource server must not be allowed to accept access tokens that are not meant for its consumption. To provide protection against token disclosure two approaches are possible, namely (a) not to include sensitive information inside the token or (b) to ensure confidentiality protection. The latter approach requires at least the communication interaction between the client and the authorization server as well as the interaction between the client and the resource server to experience confidentiality protection. As an example, TLS with a ciphersuite that offers confidentiality protection has to be applied as per . Encrypting the token content itself is another alternative. In our scenario the authorization server would, for example, encrypt the token content with a symmetric key shared with the resource server. To deal with token reuse more choices are available.

In this approach confidentiality protection of the exchange is provided on the communication interfaces between the client and the resource server, and between the client and the authorization server. No eavesdropper on the wire is able to observe the token exchange. Consequently, a replay by a third party is not possible. An authorization server wants to ensure that it only hands out tokens to clients it has authenticated first and who are authorized. For this purpose, authentication of the client to the authorization server will be a requirement to ensure adequate protection against a range of attacks. This is, however, true for the description in and as well. Furthermore, the client has to make sure it does not distribute (or leak) the access token to entities other than the intended the resource server. For that purpose the client will have to authenticate the resource server before transmitting the access token.

Instead of providing confidentiality protection, the authorization server could also put the identifier of the client into the protected token with the following semantic: 'This token is only valid when presented by a client with the following identifier.' When the access token is then presented to the resource server how does it know that it was provided by the client? It has to authenticate the client! There are many choices for authenticating the client to the resource server, for example by using client certificates in TLS , or pre-shared secrets within TLS . The choice of the preferred authentication mechanism and credential type may depend on a number of factors, including security properties available infrastructure library support credential cost (financial) performance integration into the existing IT infrastructure operational overhead for configuration and distribution of credentials This long list hints to the challenge of selecting at least one mandatory-to-implement client authentication mechanism.

A variation of the mechanism of sender authentication, described in , is to replace authentication with the proof-of-possession of a specific (session) key, i.e., key confirmation. In this model the resource server would not authenticate the client itself but would rather verify whether the client knows the session key associated with a specific access token. Examples of this approach can be found with the OAuth 1.0 MAC token , and Kerberos when utilizing the AP_REQ/AP_REP exchange (see also for a comparison between Kerberos and OAuth). To illustrate key confirmation, the first example is borrowed from Kerberos and use symmetric key cryptography. Assume that the authorization server shares a long-term secret with the resource server, called K(Authorization Server-Resource Server). This secret would be established between them out-of-band. When the client requests an access token the authorization server creates a fresh and unique session key Ks and places it into the token encrypted with the long term key K(Authorization Server-Resource Server). Additionally, the authorization server attaches Ks to the response message to the client (in addition to the access token itself) over a confidentiality protected channel. When the client sends a request to the resource server it has to use Ks to compute a keyed message digest for the request (in whatever form or whatever layer). The resource server, when receiving the message, retrieves the access token, verifies it and extracts K(Authorization Server-Resource Server) to obtain Ks. This key Ks is then used to verify the keyed message digest of the request message. Note that in this example one could imagine that the mechanism to protect the token itself is based on a symmetric key based mechanism to avoid any form of public key infrastructure but this aspect is not further elaborated in the scenario. A similar mechanism can also be designed using asymmetric cryptography. When the client requests an access token the authorization server creates an ephemeral public / privacy key pair (PK/SK) and places the public key PK into the protected token. When the authorization server returns the access token to the client it also provides the PK/SK key pair over a confidentiality protected channel. When the client sends a request to the resource server it has to use the privacy key SK to sign the request. The resource server, when receiving the message, retrieves the access token, verifies it and extracts the public key PK. It uses this ephemeral public key to verify the attached signature.

As a high level message, there are various ways the threats can be mitigated. While the details of each solution are somewhat different, they all accomplish the goal of mitigating the threats. The three approaches are: The weak point with this approach, which is briefly described in , is that the client has to be careful to whom it discloses the access token. What can be done with the token entirely depends on what rights the token entitles the presenter and what constraints it contains. A token could encode the identifier of the client but there are scenarios where the client is not authenticated to the resource server or where the identifier of the client rather represents an application class rather than a single application instance. As such, it is possible that certain deployments choose a rather liberal approach to security and that everyone who is in possession of the access token is granted access to the data. The weak point with this approach, which is briefly described in , is to setup the authentication infrastructure such that clients can be authenticated towards resource servers. Additionally, the authorization server must encode the identifier of the client in the token for later verification by the resource server. Depending on the chosen layer for providing client-side authentication there may be additional challenges due to Web server load balancing, lack of API access to identity information, etc. The weak point with this approach, see , is the increased complexity: a complete key distribution protocol has to be defined. In all cases above it has to be ensured that the client is able to keep the credentials secret.

The proof-of-possession security concept assumes that the authorization server acts as a trusted third party that binds keys to access tokens. These keys are then used by the client to demonstrate the possession of the secret to the resource server when accessing the resource. The resource server, when receiving an access token, needs to verify that the key used by the client matches the one included in the access token. There are slight differences between the use of symmetric keys and asymmetric keys when they are bound to the access token and the subsequent interaction between the client and the authorization server when demonstrating possession of these keys. shows the symmetric key procedure and illustrates how asymmetric keys are used. While symmetric cryptography provides better performance properties the use of asymmetric cryptography allows the client to keep the private key locally and never expose it to any other party. For example, with the JSON Web Token (JWT) a standardized format for access tokens is available. The necessary elements to bind symmetric or asymmetric keys to a JWT are described in . Note: The negotiation of cryptographic algorithms between the client and the authorization server is not shown in the examples below and assumed to be present in a protocol solution to meet the requirements for crypto-agility.

+---------------+ ^| | // | Authorization | / | Server | // | | / | | (I) // /+---------------+ Access / // Token / / Request // // (II) Access Token +Params / / +Symmetric Key // // / v +-----------+ +------------+ | | | | | | | Resource | | Client | | Server | | | | | | | | | +-----------+ +------------+ In order to request an access token the client interacts with the authorization server as part of the a normal grant exchange, as shown in . However, it needs to include additional information elements for use with the PoP security mechanism, as depicted in message (I). In message (II) the authorization server then returns the requested access token. In addition to the access token itself, the symmetric key is communicated to the client. This symmetric key is a unique and fresh session key with sufficient entropy for the given lifetime. Furthermore, information within the access token ties it to this specific symmetric key. Note: For this security mechanism to work the client as well as the resource server need to have access to the session key. While the key transport mechanism from the authorization server to the client has been explained in the previous paragraph there are three ways for communicating this session key from the authorization server to the resource server, namely Embedding the symmetric key inside the access token itself. This requires that the symmetric key is confidentiality protected. The resource server queries the authorization server for the symmetric key. This is an approach envisioned by the token introspection endpoint . The authorization server and the resource server both have access to the same back-end database. Smaller, tightly coupled systems might prefer such a deployment strategy.

+---------------+ ^| | Access Token Req. // | Authorization | +Parameters / | Server | +[Fingerprint] // | | / | | (I) // /+---------------+ / // / / (II) // // Access Token / / +[ephemeral // // asymmetric key pair] / v +-----------+ +------------+ | | | | | | | Resource | | Client | | Server | | | | | | | | | +-----------+ +------------+ The use of asymmetric keys is slightly different since the client or the server could be involved in the generation of the ephemeral key pair. This exchange is shown in . If the client generates the key pair it either includes a fingerprint of the public key or the public key in the request to the authorization server. The authorization server would include this fingerprint or public key in the confirmation claim inside the access token and thereby bind the asymmetric key pair to the token. If the client did not provide a fingerprint or a public key in the request then the authorization server is asked to create an ephemeral asymmetric key pair, binds the fingerprint of the public key to the access token, and returns the asymmetric key pair (public and private key) to the client. Note that there is a strong preference for generating the private/public key pair locally at the client rather than at the server.

The specification describing the interaction between the client and the authorization server, as shown in and in , can be found in . Once the client has obtained the necessary access token and keying material it can start to interact with the resource server. To demonstrate possession of the key bound to the access token it needs to apply this key to the request by computing a keyed message digest (i.e., a symmetric key-based cryptographic primitive) or a digital signature (i.e., an asymmetric cryptographic computation). When the resource server receives the request it verifies it and decides whether access to the protected resource can be granted. This exchange is shown in .

+---------------+ | | | Authorization | | Server | | | | | +---------------+ Request +-----------+ + Signature/MAC (a) +------------+ | |---------------------->| | | | [+Access Token] | Resource | | Client | | Server | | | Response (b) | | | |<----------------------| | +-----------+ [+ Signature/MAC] +------------+ ^ ^ | | | | Symmetric Key Symmetric Key or or Asymmetric Key Pair Public Key (Client) + + Parameters Parameters The specification describing the ability to sign the HTTP request from the client to the resource server can be found in .

So far the examples talked about access tokens that are passed by value and allow the resource server to make authorization decisions immediately after verifying the request from the client. In some deployments a real-time interaction between the authorization server and the resource server is envisioned that lowers the need to pass self-contained access tokens around. In that case the access token merely serves as a handle or a reference to state stored at the authorization server. As a consequence, the resource server cannot autonomously make an authorization decision when receiving a request from a client but has to consult the authorization server. This can, for example, be done using the token introspection endpoint (see ). shows the protocol interaction graphically. Despite the additional token exchange previous descriptions about associating symmetric and asymmetric keys to the access token are still applicable to this scenario.

+---------------+ Access ^| | Token Req. // | Authorization |^ (I) / | Server | \ (IV) Token // | | \ Introspection Req. / | | \ +Access // /+---------------+ \ Token / // (II) \ \\ / / Access \ \ // // Token \ (V) \ / / \Resp.\ // // \ \ / v V \ +-----------+ Request +Signature/MAC+------------+ | | (III) +Access Token | | | |---------------------->| Resource | | Client | (VI) Success or | Server | | | Failure | | | |<----------------------| | +-----------+ +------------+

The purpose of this document is to provide use cases, requirements, and motivation for developing an OAuth security solution extending Bearer Tokens. As such, this document is only about security.

This document does not require actions by IANA.

This document is the result of conference calls late 2012/early 2013 and in design team conference calls February 2013 of the IETF OAuth working group. The following persons (in addition to the OAuth WG chairs, Hannes Tschofenig, and Derek Atkins) provided their input during these calls: Bill Mills, Justin Richer, Phil Hunt, Prateek Mishra, Mike Jones, George Fletcher, Leif Johansson, Lucy Lynch, John Bradley, Tony Nadalin, Klaas Wierenga, Thomas Hardjono, Brian Campbell In the appendix of this document we reuse content from and the authors would like thank Russ Housely and Bernard Aboba for their work on RFC 4962. We would like to thank Reddy Tirumaleswar for his review.