| Network Working Group | D. Balfanz |
| Internet-Draft | R. Hamilton |
| Expires: December 31, 2013 | Google Inc |
| June 29, 2013 |
Transport Layer Security (TLS) Channel IDs
draft-balfanz-tls-channelid-01
This document describes a Transport Layer Security (TLS) extension for identifying client machines at the TLS layer without using bearer tokens.
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at http://datatracker.ietf.org/drafts/current/.
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."
This Internet-Draft will expire on December 31, 2013.
Copyright (c) 2013 IETF Trust and the persons identified as the document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (http://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.
Many applications on the Internet use bearer tokens to authenticate clients to servers. The most prominent example is the HTTP-based World Wide Web, which overwhelmingly uses HTTP cookies to authenticate client requests. Other examples include OpenID or SAML assertions, and OAuth tokens. All these have in common that the bearer of the HTTP cookie or authentication token is granted access to a protected resource, regardless of the channel over which the token is presented, or who presented it.
As a result, an adversary that manages to steal a bearer token from a client can impersonate that client to services that require the token.
This document describes a light-weight mechanism for establishing a cryptographic channel between client and server. A server can choose to bind authentication tokens to this channel, thus rendering the theft of authentication tokens fruitless - tokens must be sent over the channel to which they are bound (i.e., by the client to which they were issued) or else they will be ignored.
This document does not prescribe how authentication tokens are bound to the underlying channel. Rather, it prescribes how a client can establish a long-lived channel with a server. Such a channel persists across HTTP requests, TLS connections, and even multiple TLS sessions, as long as the same client communicates with the same server.
The basic idea is that the client proves, during the TLS handshake, possession of a private key. The corresponding public key becomes the "Channel ID" that identifies this TLS connection. Clients should re-use the same private/public key pair across subsequent TLS connections to the same server, thus creating TLS connections that share the same Channel ID.
Using private/public key pairs to define a channel (as opposed to, say, an HTTP session cookie) has several advantages: One, the credential establishing the channel (the private key) is never sent from client to server, thus removing it from the reach of eavesdroppers in the network. Two, clients can choose to implement cryptographic operations in a secure hardware module, which further removes the private key from the reach of eavesdroppers residing on the client itself.
TLS already supports a means of identifying clients without using bearer tokens: client certificates. However, a number of problems with using client certificates motivated the development of an alternative.
Most importantly, it's not acceptable for a client identifier to be transmitted in the clear, because eavesdroppers in the network could use these identifiers to deanonymize TLS connections. Client certificates in TLS, however, are sent unencrypted. Although we could also define a change to the TLS state machine to move the client certificates under encryption, such changes eliminate most of the benefits of reusing something that's already defined.
TLS client certificates are also defined to be part of the session state. Even though the key material used for TLS client authentication might be protected from theft from compromised clients (for example, by employing hardware secure elements on the client), TLS session resumption information rarely is. Because client certificates are part of the session state, stolen session resumption information gives the attacker something equivalent to a stolen client private key. Our objective, however, is that attackers should not be able to give the impression that they can wield a private key unless they are actually in control of that private key.
Client-certificates typically identify a user, while we seek to identify machines. Since they are not, conceptually, mutually exclusive and as only a single client certificate can be provided in TLS, we don't want to consume that single slot and eliminate the possibility of also using existing client certificates.
Client certificates are implemented in TLS as X.509 certificates and we don't wish to require servers to parse arbitrary ASN.1. ASN.1 is a complex encoding that has been the source of several security vulnerabilities in the past and typical TLS servers can currently avoid doing ASN.1 parsing.
X.509 certificates always include a signature, which would be a self-signature in this case. Calculating and transmitting the self-signature is a waste of computation and network traffic in our use. Although we could define a null signature algorithm with an empty signature, such deviations from X.509 eliminate many of the benefits of reusing something that is already implemented.
Finally, client certificates trigger significant server-side processing by default and often need to be stored in their entirety for the duration of the connection. Since this design is intended to be widely used, it allows servers to retain only a cryptographic hash of the client's public key after the handshake completes.
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119 [RFC2119].
For the purpose of this specification, a public key is a point Q = dG on the P-256 curve [DSS] (where d is the ECC private key, and G is the curve base point). Clients SHOULD use a separate key pair (d, Q) for each server they connect to, and generate a new key pair if necessary according to appendix B.4 in FIPS-186-3 [DSS].
A public key Q has two affine coordinates x, y: Q = (x,y). The public key Q - or, in other words, the pair x, y - that a client uses for a specific server is that client's Channel ID for that server.
A new extension type (channel_id(TBD)) is defined and MAY be included by the client in its ClientHello message. If, and only if, the server sees this extension in the ClientHello, it MAY choose to echo the extension in its ServerHello. In both cases, the extension_data field MUST be empty.
enum {
channel_id(TBD), (65535)
} ExtensionType;
A new handshake message type (encrypted_extensions(TBD)) is defined. If the server included a channel_id extension in its ServerHello message, the client MUST verify that the selected cipher suite is sufficiently strong. If the cipher suite provides < 80-bits of security, the client MUST abort the handshake with a fatal illegal_parameter alert. Otherwise, the client MUST send an EncryptedExtensions message after its ChangeCipherSpec and before its Finished message.
enum {
encrypted_extensions(TBD), (65535)
} HandshakeType;
Therefore a full handshake with EncryptedExtensions has the following flow (contrast with section 7.3 of RFC 5246 [RFC5246]):
Client Server
ClientHello (ChannelID extension) -------->
ServerHello
(ChannelID extension)
Certificate*
ServerKeyExchange*
CertificateRequest*
<-------- ServerHelloDone
Certificate*
ClientKeyExchange
CertificateVerify*
[ChangeCipherSpec]
EncryptedExtensions
Finished -------->
[ChangeCipherSpec]
<-------- Finished
Application Data <-------> Application Data
An abbreviated handshake with EncryptedExtensions has the following flow:
Client Server
ClientHello (ChannelID extension) -------->
ServerHello
(ChannelID extension)
[ChangeCipherSpec]
<-------- Finished
[ChangeCipherSpec]
EncryptedExtensions
Finished -------->
Application Data <-------> Application Data
The EncryptedExtensions message contains a series of Extension structures (see section 7.4.1.4 of RFC 5246 [RFC5246]
If the server included a channel_id extension in its ServerHello message, the client MUST include, within an EncryptedExtensions message, an Extension with extension_type equal to channel_id(TBD). The extension_data of which has the following format:
struct {
opaque x[32];
opaque y[32];
opaque r[32];
opaque s[32];
} ChannelIDExtension;
The contents of each of x, y, r and s is a 32-byte, big-endian number. The x and y fields contain the affine coordinates of the client's Channel ID Q (i.e., a P-256 [DSS] curve point). The r and s fields contain an ECDSA [DSS] signature by the corresponding private key over this US-ASCII strong (not including quotes, and where "\x00" represents an octet containing all zero bits):
"TLS Channel ID signature\x00"
followed by hashes of both the client-sent and server-sent handshake messages, as seen by the client, prior to the EncryptedExtensions message.
Unlike many other TLS extensions, this extension does not establish properties of the session, only of the connection. When session resumption or session tickets [RFC5077] are used, the previous contents of this extension are irrelevant and only the values in the new handshake messages are considered.
There are four classes of attackers against which we consider our security guarantees: passive network attackers, active network attackers, active network attackers with misissued certificates and attackers in possession of the legitimate server's private key.
First, we wish to guarantee that we don't disclose the Channel ID to passive or active network attackers. We do this by sending a constant-length Channel ID under encryption. However, since the Channel ID may be transmitted before the server's Finished message is received, it's possible that the server isn't in possession of the corresponding private key to the certificate that it presented. In this situation, an active attacker could cause a Channel ID to be transmitted under a random key in a cipher suite of their choosing. Therefore we limit the permissible cipher suites to those where decrypting the message is infeasible.
Even with this limit, an active attacker can cause the Channel ID to be transmitted in a non-forward-secure manner. Subsequent disclosure of the server's private key would allow previously recorded Channel IDs to be decrypted.
Second, we wish to guarantee that none of the first three attackers can terminate/hijack a TLS connection and impersonate a Channel ID from that connection when connecting to the legitimate server. We assume that TLS provides sufficient security to prevent these attackers from being able to hijack the TLS connection. An active attacker illegitimately in possession of a certificate for a server can successfully terminate a TLS connection destined for that server and decrypt the Channel ID. However, as the signature covers the handshake hashes, and therefore the server's certificate, it wouldn't be accepted by the true server.
Against an attacker with the legitimate server's private key we can provide the second guarantee only if the legitimate server uses a forward-secret cipher suite, otherwise the attacker can hijack the connection.
An HTTP application on the server can channel-bind its cookies by associating them with the Channel ID of the user-agent that the cookies are being set on. The server MAY then choose to consider cookies sent from the user-agent invalid if the Channel ID associated with the cookie does not match the Channel ID used by the user-agent when it sends the cookie back to the server.
Such a mismatch could occur when the cookie has been obtained from the legitimate user-agent and is now being sent by a client not in possession of the legitimate user-agent's Channel ID private key. The mismatch can also occur if the legitimate user-agent has changed the Channel ID it is using for the server, presumably due to the user requesting a Channel ID reset through the user-agent's user interface (see Section 8). Such a user intervention is analogous to the user's removal of cookies from the user-agent, but instead of removing cookies, the cookies are being rendered invalid (in the eyes of the server).
Similarly to cookies, a server may choose to channel-bind OAuth tokens (or any other kind of authorization tokens) to the clients to which they are issued. The mechanism on the server remains the same (it associates the OAuth token with the client's Channel ID either by storing this information in a database, or by suitably encoding the information in the OAuth token itself), but the application-level protocol may be different: In addition to HTTP, OAuth tokens are used in protocols such as IMAP and XMPP.
The TLS layer does its part in protecting user privacy by transmitting the Channel ID public key under encryption. Higher levels of the stack must ensure that the same Channel ID is not used with different servers in such a way as to provide a linkable identifier. For example, a user-agent must use different Channel IDs for communicating with different servers. Because channel-bound cookies are an important use case for TLS Channel ID, and cookies can be set on top-level domains, it is RECOMMENDED that user-agents use the same Channel ID for servers within the same top-level domain, and different Channel IDs for different top-level domains. User-agents must also ensure that Channel ID state can be reset by the user in the same way as other identifiers, i.e. cookies.
However, there are some security concerns that could result in the disclosure of a client's Channel ID to a network attacker. This is covered in the Security Considerations section.
Clients that share an IP address can be disambiguated through their Channel IDs. This is analogous to protocols that use cookies (e.g., HTTP), which also allow disambiguation of user-agents behind proxies.
Channel ID has been designed to provide privacy equivalent to that of cookies. User-agents SHOULD continue to meet this design goal at higher layers of the protocol stack. For example, if a user indicates that they would like to block third-party cookies (or if the user-agent has some sort of policy around when it blocks third-party cookies by default), then the user agent SHOULD NOT use Channel ID on third-party connections (or other connections through which the user-agent would refuse to send or accept cookies).
This document requires IANA to update its registry of TLS extensions to assign an entry referred to here as channel_id.
This document also requires IANA to update its registry of TLS handshake types to assign an entry referred to here as encrypted_extensions.
| [RFC2119] | Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, March 1997. |
| [RFC5246] | Dierks, T. and E. Rescorla, "The Transport Layer Security (TLS) Protocol Version 1.2", RFC 5246, August 2008. |
| [DSS] | National Institute of Standards and Technology, "FIPS 186-3: Digital Signature Standard", |
| [RFC5077] | Salowey, J., Zhou, H., Eronen, P. and H. Tschofenig, "Transport Layer Security (TLS) Session Resumption without Server-Side State", RFC 5077, January 2008. |
The following individuals contributed to this specification:
Dirk Balfanz, Wan-Teh Chang, Ryan Hamilton, Adam Langley, and Mayank Upadhyay.