Secondary Certificate Authentication in HTTP/2

HTTP clients need to know that the content they receive on a connection comes from the origin that they intended to retrieve in from. The traditional form of server authentication in HTTP has been in the form of X.509 certificates provided during the TLS RFC5246 handshake. Many existing HTTP servers also have authentication requirements for the resources they serve. Of the bountiful authentication options available for authenticating HTTP requests, client certificates present a unique challenge for resource-specific authentication requirements because of the interaction with the underlying TLS layer. TLS 1.2 supports one server and one client certificate on a connection. These certificates may contain multiple identities, but only one certificate may be provided.

Section 9.1.1 of describes how connections may be used to make requests from multiple origins as long as the server is authoritative for both. A server is considered authoritative for an origin if DNS resolves the origin to the IP address of the server and (for TLS) if the certificate presented by the server contains the origin in the Subject Alternative Names field. enables a step of abstraction from the DNS resolution. If both hosts have provided an Alternative Service at hostnames which resolve to the IP address of the server, they are considered authoritative just as if DNS resolved the origin itself to that address. However, the server’s one TLS certificate is still required to contain the name of each origin in question. relaxes the requirement to perform the DNS lookup if already connected to a server with an appropriate certificate which claims support for a particular origin. Servers which host many origins often would prefer to have separate certificates for some sets of origins. This may be for ease of certificate management (the ability to separately revoke or renew them), due to different sources of certificates (a CDN acting on behalf of multiple origins), or other factors which might drive this administrative decision. Clients connecting to such origins cannot currently reuse connections, even if both client and server would prefer to do so. Because the TLS SNI extension is exchanged in the clear, clients might also prefer to retrieve certificates inside the encrypted context. When this information is sensitive, it might be advantageous to request a general-purpose certificate or anonymous ciphersuite at the TLS layer, while acquiring the “real” certificate in HTTP after the connection is established.

For servers that wish to use client certificates to authenticate users, they might request client authentication during or immediately after the TLS handshake. However, if not all users or resources need certificate-based authentication, a request for a certificate has the unfortunate consequence of triggering the client to seek a certificate, possibly requiring user interaction, network traffic, or other time-consuming activities. During this time, the connection is stalled in many implementations. Such a request can result in a poor experience, particularly when sent to a client that does not expect the request. The TLS 1.3 CertificateRequest can be used by servers to give clients hints about which certificate to offer. Servers that rely on certificate-based authentication might request different certificates for different resources. Such a server cannot use contextual information about the resource to construct an appropriate TLS CertificateRequest message during the initial handshake. Consequently, client certificates are requested at connection establishment time only in cases where all clients are expected or required to have a single certificate that is used for all resources. Many other uses for client certificates are reactive, that is, certificates are requested in response to the client making a request.

In HTTP/1.1, a server that relies on client authentication for a subset of users or resources does not request a certificate when the connection is established. Instead, it only requests a client certificate when a request is made to a resource that requires a certificate. TLS 1.2 accomodates this by permitting the server to request a new TLS handshake, in which the server will request the client’s certificate. shows the server initiating a TLS-layer renegotiation in response to receiving an HTTP/1.1 request to a protected resource.

*1 <---------------------- (TLS) HelloRequest -- *2 -- (TLS) ClientHello -----------------------> <------------------ (TLS) ServerHello, ... -- <---------------- (TLS) CertificateRequest -- *3 -- (TLS) ..., Certificate ------------------> *4 -- (TLS) Finished --------------------------> <-------------------------- (TLS) Finished -- <--------------------------- (HTTP) 200 OK -- *5 ]]> In this example, the server receives a request for a protected resource (at *1 on ). Upon performing an authorization check, the server determines that the request requires authentication using a client certificate and that no such certificate has been provided. The server initiates TLS renegotiation by sending a TLS HelloRequest (at *2). The client then initiates a TLS handshake. Note that some TLS messages are elided from the figure for the sake of brevity. The critical messages for this example are the server requesting a certificate with a TLS CertificateRequest (*3); this request might use information about the request or resource. The client then provides a certificate and proof of possession of the private key in Certificate and CertificateVerify messages (*4). When the handshake completes, the server performs any authorization checks a second time. With the client certificate available, it then authorizes the request and provides a response (*5).

TLS 1.3 introduces a new client authentication mechanism that allows for clients to authenticate after the handshake has been completed. For the purposes of authenticating an HTTP request, this is functionally equivalent to renegotiation. shows the simpler exchange this enables.

<---------------- (TLS) CertificateRequest -- -- (TLS) Certificate, CertificateVerify, Finished -----------------------> <--------------------------- (HTTP) 200 OK -- ]]> TLS 1.3 does not support renegotiation, instead supporting direct client authentication. In contrast to the TLS 1.2 example, in TLS 1.3, a server can simply request a certificate.

An important part of the HTTP/1.1 exchange is that the client is able to easily identify the request that caused the TLS renegotiation. The client is able to assume that the next unanswered request on the connection is responsible. The HTTP stack in the client is then able to direct the certificate request to the application or component that initiated that request. This ensures that the application has the right contextual information for processing the request. In HTTP/2, a client can have multiple outstanding requests. Without some sort of correlation information, a client is unable to identify which request caused the server to request a certificate. Thus, the minimum necessary mechanism to support reactive certificate authentication in HTTP/2 is an identifier that can be use to correlate an HTTP request with a request for a certificate. Since streams are used for individual requests, correlation with a stream is sufficient. prohibits renegotiation after any application data has been sent. This completely blocks reactive certificate authentication in HTTP/2 using TLS 1.2. If this restriction were relaxed by an extension or update to HTTP/2, such an identifier could be added to TLS 1.2 by means of an extension to TLS. Unfortunately, many TLS 1.2 implementations do not permit application data to continue during a renegotiation. This is problematic for a multiplexed protocol like HTTP/2.

This draft defines HTTP/2 frames to carry the relevant certificate messages, enabling certificate-based authentication of both clients and servers independent of TLS version. This mechanism can be implemented at the HTTP layer without breaking the existing interface between HTTP and applications above it. This could be done in a naive manner by replicating the TLS messages as HTTP/2 frames on each stream. However, this would create needless redundancy between streams and require frequent expensive signing operations. Instead, TLS Exported Authenticators are exchanged on stream zero and the on-stream frames incorporate them by reference as needed. TLS Exported Authenticators are structured messages that can be exported by either party of a TLS connection and validated by the other party. An authenticator message can be constructed by either the client or the server given an established TLS connection, a certificate, and a corresponding private key. Exported Authenticators use the message structures from section 4.4 of , but different parameters. Each Authenticator is computed using a Handshake Context and Finished MAC Key derived from the TLS session. The Handshake Context is identical for both parties of the TLS connection, while the Finished MAC Key is dependent on whether the Authenticator is created by the client or the server. Successfully verified Authenticators result in certificate chains, with verified possession of the corresponding private key, which can be supplied into a collection of available certificates. Likewise, descriptions of desired certificates can be supplied into these collections. These pre-supplied elements are then available for automatic use (in some situations) or for reference by individual streams. describes how the feature is employed, defining means to detect support in peers (), make certificates and requests available (), and indicate when streams are blocked waiting on an appropriate certificate (). defines the required frame types, which parallel the TLS 1.3 message exchange. Finally, defines new error types which can be used to notify peers when the exchange has not been successful.

RFC 2119 defines the terms “MUST”, “MUST NOT”, “SHOULD” and “MAY”.

A certificate chain with proof of possession of the private key corresponding to the end-entity certificate is sent as a single CERTIFICATE frame (see ) on stream zero. Once the holder of a certificate has sent the chain and proof, this certificate chain is cached by the recipient and available for future use. If the certificate is marked as AUTOMATIC_USE, the certificate may be used by the recipient to authorize any current or future request. Otherwise, the recipient requests the required certificate on each stream, but the previously-supplied certificates are available for reference without having to resend them. Likewise, the details of a request are sent on stream zero and stored by the recipient. These details will be referenced by subsequent CERTIFICATE_NEEDED frames. Data sent by each peer is correlated by the ID given in each frame. This ID is unrelated to values used by the other peer, even if each uses the same ID in certain cases.

Clients and servers that will accept requests for HTTP-layer certificate authentication indicate this using the HTTP/2 SETTINGS_HTTP_CERT_AUTH (0xSETTING-TBD) setting. The initial value for the SETTINGS_HTTP_CERT_AUTH setting is 0, indicating that the peer does not support HTTP-layer certificate authentication. If a peer does support HTTP-layer certificate authentication, the value is 1.

When a peer has advertised support for HTTP-layer certificates as in , either party can supply additional certificates into the connection at any time. These certificates then become available for the peer to consider when deciding whether a connection is suitable to transport a particular request. Available certificates which have the AUTOMATIC_USE flag set MAY be used by the recipient without further notice. This means that clients or servers which predict a certificate will be required could pre-supply the certificate without being asked. Regardless of whether AUTOMATIC_USE is set, these certificates are available for reference by future USE_CERTIFICATE frames.

<----------------------- (stream N) 200 OK -- ]]>

-- (streams 1,3) GET /protected ------------> <-------------------- (streams 1,3) 200 OK -- ]]> Likewise, either party can supply a CERTIFICATE_REQUEST that outlines parameters of a certificate they might request in the future. It is important to note that this does not currently request such a certificate, but makes the contents of the request available for reference by a future CERTIFICATE_NEEDED frame.

As defined in , when a client finds that a https:// origin (or Alternative Service ) to which it needs to make a request has the same IP address as a server to which it is already connected, it MAY check whether the TLS certificate provided contains the new origin as well, and if so, reuse the connection. If the TLS certificate does not contain the new origin, but the server has claimed support for that origin (with an ORIGIN frame, see ) and advertised support for HTTP-layer certificates (see ), it MAY send a CERTIFICATE_NEEDED frame on the stream it will use to make the request. (If the request parameters have not already been made available using a CERTIFICATE_REQUEST frame, the client will need to send the CERTIFICATE_REQUEST in order to generate the CERTIFICATE_NEEDED frame.) The stream represents a pending request to that origin which is blocked until a valid certificate is processed. The request is blocked until the server has responded with a USE_CERTIFICATE frame pointing to a certificate for that origin. If the certificate is already available, the server SHOULD immediately respond with the appropriate USE_CERTIFICATE frame. (If the certificate has not already been transmitted, the server will need to make the certificate available as described in before completing the exchange.) If the server does not have the desired certificate, it MUST respond with an empty USE_CERTIFICATE frame. In this case, or if the server has not advertised support for HTTP-layer certificates, the client MUST NOT send any requests for resources in that origin on the current connection.

... -- (stream N) CERTIFICATE_NEEDED -----------> <------------------ (stream 0) CERTIFICATE -- <-------------- (stream N) USE_CERTIFICATE -- -- (stream N) GET /from-new-origin ---------> <----------------------- (stream N) 200 OK -- ]]> Likewise, on each stream where certificate authentication is required, the server sends a CERTIFICATE_NEEDED frame, which the client answers with a USE_CERTIFICATE frame indicating the certificate to use. If the request parameters or the responding certificate are not already available, they will need to be sent as described in as part of this exchange.

<----------- (stream N) CERTIFICATE_NEEDED -- -- (stream 0) CERTIFICATE ------------------> -- (stream N) USE_CERTIFICATE --------------> <----------------------- (stream N) 200 OK -- ]]> A server SHOULD provide certificates for an origin before pushing resources from it or supplying content referencing the origin. If a client receives a PUSH_PROMISE referencing an origin for which it has not yet received the server’s certificate, the client MUST verify the server’s possession of an appropriate certificate by sending a CERTIFICATE_NEEDED frame on the pushed stream to inform the server that progress is blocked until the request is satisfied. The client MUST NOT use the pushed resource until an appropriate certificate has been received and validated.

The CERTIFICATE_REQUEST and CERTIFICATE_NEEDED frames are correlated by their Request-ID field. Subsequent CERTIFICATE_NEEDED frames with the same Request-ID value MAY be sent on other streams where the sender is expecting a certificate with the same parameters. The CERTIFICATE, and USE_CERTIFICATE frames are correlated by their Cert-ID field. Subsequent USE_CERTIFICATE frames with the same Cert-ID MAY be sent in response to other CERTIFICATE_NEEDED frames and refer to the same certificate. Request-ID and Cert-ID are sender-local, and the use of the same value by the other peer does not imply any correlation between their frames. These values MUST be unique per sender over the lifetime of the connection.

The CERTIFICATE_NEEDED frame (0xFRAME-TBD1) is sent to indicate that the HTTP request on the current stream is blocked pending certificate authentication. The frame includes a request identifier which can be used to correlate the stream with a previous CERTIFICATE_REQUEST frame sent on stream zero. The CERTIFICATE_REQUEST describes the certificate the sender requires to make progress on the stream in question. The CERTIFICATE_NEEDED frame contains 2 octets, which is the authentication request identifier, Request-ID. A peer that receives a CERTIFICATE_NEEDED of any other length MUST treat this as a stream error of type PROTOCOL_ERROR. Frames with identical request identifiers refer to the same CERTIFICATE_REQUEST. A server MAY send multiple CERTIFICATE_NEEDED frames on the same stream. If a server requires that a client provide multiple certificates before authorizing a single request, each required certificate MUST be indicated with a separate CERTIFICATE_NEEDED frame, each of which MUST have a different request identifier (referencing different CERTIFICATE_REQUEST frames describing each required certificate). To reduce the risk of client confusion, servers SHOULD NOT have multiple outstanding CERTIFICATE_NEEDED frames on the same stream at any given time. Clients MUST NOT send multiple CERTIFICATE_NEEDED frames on the same stream. The CERTIFICATE_NEEDED frame MUST NOT be sent to a peer which has not advertised support for HTTP-layer certificate authentication. The CERTIFICATE_NEEDED frame MUST NOT be sent on stream zero, and MUST NOT be sent on a stream in the “half-closed (local)” state . A client that receives a CERTIFICATE_NEEDED frame on a stream which is not in a valid state SHOULD treat this as a stream error of type PROTOCOL_ERROR.

The USE_CERTIFICATE frame (0xFRAME-TBD4) is sent in response to a CERTIFICATE_NEEDED frame to indicate which certificate is being used to satisfy the requirement. A USE_CERTIFICATE frame with no payload refers to the certificate provided at the TLS layer, if any. If no certificate was provided at the TLS layer, the stream should be processed with no authentication, likely returning an authentication-related error at the HTTP level (e.g. 403) for servers or routing the request to a new connection for clients. Otherwise, the USE_CERTIFICATE frame contains the two-octet Cert-ID of the certificate the sender wishes to use. This MUST be the ID of a certificate for which proof of possession has been presented in a CERTIFICATE frame. Recipients of a USE_CERTIFICATE frame of any other length MUST treat this as a stream error of type PROTOCOL_ERROR. Frames with identical certificate identifiers refer to the same certificate chain. The USE_CERTIFICATE frame MUST NOT be sent on stream zero or a stream on which a CERTIFICATE_NEEDED frame has not been received. Receipt of a USE_CERTIFICATE frame in these circumstances SHOULD be treated as a stream error of type PROTOCOL_ERROR. Each USE_CERTIFICATE frame should reference a preceding CERTIFICATE frame. Receipt of a USE_CERTIFICATE frame before the necessary frames have been received on stream zero MUST also result in a stream error of type PROTOCOL_ERROR. The referenced certificate chain MUST conform to the requirements expressed in the CERTIFICATE_REQUEST to the best of the sender’s ability. Specifically, if the CERTIFICATE_REQUEST contained a non-empty Cert-Extensions element, the end-entity certificate MUST match with regard to the extensions recognized by the sender. If these requirements are not satisfied, the recipient MAY at its discretion either return an error at the HTTP semantic layer, or respond with a stream error on any stream where the certificate is used. defines certificate-related error codes which might be applicable.

TLS 1.3 defines the CertificateRequest message, which prompts the client to provide a certificate which conforms to certain properties specified by the server. This draft defines the CERTIFICATE_REQUEST frame (0xFRAME-TBD2), which uses the same set of extensions to specify a desired certificate, but can be sent over any TLS version and can be sent by either peer. The CERTIFICATE_REQUEST frame SHOULD NOT be sent to a peer which has not advertised support for HTTP-layer certificate authentication. The CERTIFICATE_REQUEST frame MUST be sent on stream zero. A CERTIFICATE_REQUEST frame received on any other stream MUST be rejected with a stream error of type PROTOCOL_ERROR.

The frame contains the following fields: Request-ID is a 16-bit opaque identifier used to correlate subsequent certificate-related frames with this request. The identifier MUST be unique in the session for the sender. A list of certificate selection criteria, represented in a series of Extension structures (see section 4.2). This criteria MUST be used in certificate selection as described in . The number of Extension structures is given by the 16-bit Extension-Count field, which MAY be zero. Some extensions used for certificate selection allow multiple values (e.g. oid_filters on Extended Key Usage). If the sender has included a non-empty Extensions list, the certificate MUST match all criteria specified by extensions the recipient recognizes. However, the recipient MUST ignore and skip any unrecognized certificate selection extensions. Servers MUST be able to recognize the server_name extension () at a minimum. Clients MUST always specify the desired origin using this extension, though other extensions MAY also be included.

The CERTIFICATE frame (id=0xFRAME-TBD3) provides a exported authenticator message from the TLS layer that provides a chain of certificates, associated extensions and proves possession of the private key corresponding to the end-entity certificate. The CERTIFICATE frame defines two flags: Indicates that the certificate can be used automatically on future requests. Indicates that the exported authenticator spans more than one frame.

The Exported Authenticator Fragment field contains a portion of the opaque data returned from the TLS connection exported authenticator authenticate API. See for more details on the input to this API. This opaque data is transported in zero or more CERTIFICATE frames with the TO_BE_CONTINUED flag set, followed by one CERTIFICATE frame with the TO_BE_CONTINUED flag unset. Each of these frames contains the same Cert-ID field, permitting them to be associated with each other. Receipt of any CERTIFICATE frame with the same Cert-ID following the receipt of a CERTIFICATE frame with TO_BE_CONTINUED unset MUST be treated as a connection error of type PROTOCOL_ERROR. If the AUTOMATIC_USE flag is set, the recipient MAY omit sending CERTIFICATE_NEEDED frames on future streams which would require a similar certificate and use the referenced certificate for authentication without further notice to the holder. This behavior is optional, and receipt of a CERTIFICATE_NEEDED frame does not imply that previously-presented certificates were unacceptable, even if AUTOMATIC_USE was set. Servers MUST set the AUTOMATIC_USE flag when sending a CERTIFICATE frame. A server MUST NOT send certificates for origins which it is not prepared to service on the current connection. Upon receiving a complete series of CERTIFICATE frames, the receiver may validate the Exported Authenticator value by using the exported authenticator API. This returns either an error indicating that the message was invalid, or the certificate chain and extensions used to create the message. The CERTIFICATE frame MUST be sent on stream zero. A CERTIFICATE frame received on any other stream MUST be rejected with a stream error of type PROTOCOL_ERROR.

The Exported Authenticator API defined in takes as input a certificate, supporting information about the certificate (OCSP, SCT, etc.), and an optional certificate_request_context. When generating exported authenticators for use with this extension, the certificate_request_context MUST be the two-octet Cert-ID. Upon receipt of a completed authenticator, an endpoint MUST check that: the validate API confirms the validity of the authenticator itself the certificate_request_context matches the Cert-ID of the frame(s) in which it was received Once the authenticator is accepted, the endpoint can perform any other checks for the acceptability of the certificate itself.

Because this draft permits certificates to be exchanged at the HTTP framing layer instead of the TLS layer, several certificate-related errors which are defined at the TLS layer might now occur at the HTTP framing layer. In this section, those errors are restated and added to the HTTP/2 error code registry. A certificate was corrupt, contained signatures that did not verify correctly, etc. A certificate was of an unsupported type or did not contain required extensions A certificate was revoked by its signer A certificate has expired or is not currently valid Any other certificate-related error As described in , implementations MAY choose to treat a stream error as a connection error at any time. Of particular note, a stream error cannot occur on stream 0, which means that implementations cannot send non-session errors in response to CERTIFICATE_REQUEST, and CERTIFICATE frames. Implementations which do not wish to terminate the connection MAY either send relevant errors on any stream which references the failing certificate in question or process the requests as unauthenticated and provide error information at the HTTP semantic layer.

This mechanism defines an alternate way to obtain server and client certificates other than in the initial TLS handshake. While the signature of exported authenticator values is expected to be equally secure, it is important to recognize that a vulnerability in this code path is at least equal to a vulnerability in the TLS handshake.

This mechanism could increase the impact of a key compromise. Rather than needing to subvert DNS or IP routing in order to use a compromised certificate, a malicious server now only needs a client to connect to some HTTPS site under its control in order to present the compromised certificate. As recommended in , clients opting not to consult DNS ought to employ some alternative means to increase confidence that the certificate is legitimate. As noted in the Security Considerations of , it difficult to formally prove that an endpoint is jointly authoritative over multiple certificates, rather than individually authoritative on each certificate. As a result, clients MUST NOT assume that because one origin was previously colocated with another, those origins will be reachable via the same endpoints in the future. Clients MUST NOT consider previous secondary certificates to be validated after TLS session resumption. However, clients MAY proactively query for previously-presented secondary certificates.

This draft defines a mechanism which could be used to probe servers for origins they support, but opens no new attack versus making repeat TLS connections with different SNI values. Servers SHOULD impose similar denial-of-service mitigations (e.g. request rate limits) to CERTIFICATE_REQUEST frames as to new TLS connections. While the extensions in the CERTIFICATE_REQUEST frame permit the sender to enumerate the acceptable Certificate Authorities for the requested certificate, it might not be prudent (either for security or data consumption) to include the full list of trusted Certificate Authorities in every request. Senders, particularly clients, SHOULD send only the extensions that narrowly specify which certificates would be acceptable.

Failure to provide a certificate on a stream after receiving CERTIFICATE_NEEDED blocks processing, and SHOULD be subject to standard timeouts used to guard against unresponsive peers. Validating a multitude of signatures can be computationally expensive, while generating an invalid signature is computationally cheap. Implementations will require checks for attacks from this direction. Invalid exported authenticators SHOULD be treated as a session error, to avoid further attacks from the peer, though an implementation MAY instead disable HTTP-layer certificates for the current connection instead.

Implementations need to be aware of the potential for confusion about the state of a connection. The presence or absence of a validated certificate can change during the processing of a request, potentially multiple times, as USE_CERTIFICATE frames are received. A server that uses certificate authentication needs to be prepared to reevaluate the authorization state of a request as the set of certificates changes. Client implementations need to carefully consider the impact of setting the AUTOMATIC_USE flag. This flag is a performance optimization, permitting the client to avoid a round-trip on each request where the server checks for certificate authentication. However, once this flag has been sent, the client has zero knowledge about whether the server will use the referenced cert for any future request, or even for an existing request which has not yet completed. Clients MUST NOT set this flag on any certificate which is not appropriate for currently-in-flight requests, and MUST NOT make any future requests on the same connection which they are not willing to have associated with the provided certificate.

This draft adds entries in three registries. The HTTP/2 SETTINGS_HTTP_CERT_AUTH setting is registered in . Four frame types are registered in . Six error codes are registered in .

The SETTINGS_HTTP_CERT_AUTH setting is registered in the “HTTP/2 Settings” registry established in . SETTINGS_HTTP_CERT_AUTH 0xSETTING-TBD 0 This document.

Four new frame types are registered in the “HTTP/2 Frame Types” registry established in . The entries in the following table are registered by this document. Frame Type Code Specification CERTIFICATE_NEEDED 0xFRAME-TBD1 CERTIFICATE_REQUEST 0xFRAME-TBD2 CERTIFICATE 0xFRAME-TBD3 USE_CERTIFICATE 0xFRAME-TBD4

Five new error codes are registered in the “HTTP/2 Error Code” registry established in . The entries in the following table are registered by this document. Name Code Specification BAD_CERTIFICATE 0xERROR-TBD1 UNSUPPORTED_CERTIFICATE 0xERROR-TBD2 CERTIFICATE_REVOKED 0xERROR-TBD3 CERTIFICATE_EXPIRED 0xERROR-TBD4 CERTIFICATE_GENERAL 0xERROR-TBD5

Eric Rescorla pointed out several failings in an earlier revision. Andrei Popov contributed to the TLS considerations.