| Network Working Group | J. Yasskin |
| Internet-Draft | |
| Intended status: Standards Track | March 05, 2018 |
| Expires: September 6, 2018 |
Signed HTTP Exchanges
draft-yasskin-http-origin-signed-responses-03
This document specifies how a server can send an HTTP request/response pair, known as an exchange, with signatures that vouch for that exchange’s authenticity. These signatures can be verified against an origin’s certificate to establish that the exchange is authoritative for an origin even if it was transferred over a connection that isn’t. The signatures can also be used in other ways described in the appendices.
These signatures contain countermeasures against downgrade and protocol-confusion attacks.
Discussion of this draft takes place on the HTTP working group mailing list (ietf-http-wg@w3.org), which is archived at https://lists.w3.org/Archives/Public/ietf-http-wg/.
The source code and issues list for this draft can be found in https://github.com/WICG/webpackage.
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 https://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 September 6, 2018.
Copyright (c) 2018 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 (https://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.
Signed HTTP exchanges provide a way to prove the authenticity of a resource in cases where the transport layer isn’t sufficient. This can be used in several ways:
Subsequent work toward the use cases in [I-D.yasskin-webpackage-use-cases] will provide a way to group signed exchanges into bundles that can be transmitted and stored together, but single signed exchanges are useful enough to standardize on their own.
The key words “MUST”, “MUST NOT”, “REQUIRED”, “SHALL”, “SHALL NOT”, “SHOULD”, “SHOULD NOT”, “RECOMMENDED”, “NOT RECOMMENDED”, “MAY”, and “OPTIONAL” in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.
In the response of an HTTP exchange the server MAY include a Signature header field (Section 3.1) holding a list of one or more parameterised signatures that vouch for the content of the exchange. Exactly which content the signature vouches for can depend on how the exchange is transferred (Section 5).
The client categorizes each signature as “valid” or “invalid” by validating that signature with its certificate or public key and other metadata against the exchange’s headers and content (Section 3.5). This validity then informs higher-level protocols.
Each signature is parameterised with information to let a client fetch assurance that a signed exchange is still valid, in the face of revoked certificates and newly-discovered vulnerabilities. This assurance can be bundled back into the signed exchange and forwarded to another client, which won’t have to re-fetch this validity information for some period of time.
The Signature header field conveys a list of signatures for an exchange, each one accompanied by information about how to determine the authority of and refresh that signature. Each signature directly signs the exchange’s headers and identifies one of those headers that enforces the integrity of the exchange’s payload.
The Signature header is a Structured Header as defined by [I-D.ietf-httpbis-header-structure]. Its value MUST be a list (Section 4.8 of [I-D.ietf-httpbis-header-structure]) of parameterised labels (Section 4.4 of [I-D.ietf-httpbis-header-structure]).
Each parameterised label MUST have parameters named “sig”, “integrity”, “validityUrl”, “date”, and “expires”. Each parameterised label MUST also have either “certUrl” and “certSha256” parameters or an “ed25519Key” parameter. This specification gives no meaning to the label itself, which can be used as a human-readable identifier for the signature (see Section 3.1.2, Paragraph 1). The present parameters MUST have the following values:
The “certUrl” parameter is not signed, so intermediates can update it with a pointer to a cached version.
The following header is included in the response for an exchange with effective request URI https://example.com/resource.html. Newlines are added for readability.
Signature: sig1; sig=*MEUCIQDXlI2gN3RNBlgFiuRNFpZXcDIaUpX6HIEwcZEc0cZYLAIga9DsVOMM+g5YpwEBdGW3sS+bvnmAJJiSMwhuBdqp5UY; integrity="mi"; validityUrl="https://example.com/resource.validity.1511128380"; certUrl="https://example.com/oldcerts"; certSha256=*W7uB969dFW3Mb5ZefPS9Tq5ZbH5iSmOILpjv2qEArmI; date=1511128380; expires=1511733180, sig2; sig=*MEQCIGjZRqTRf9iKNkGFyzRMTFgwf/BrY2ZNIP/dykhUV0aYAiBTXg+8wujoT4n/W+cNgb7pGqQvIUGYZ8u8HZJ5YH26Qg; integrity="mi"; validityUrl="https://example.com/resource.validity.1511128380"; certUrl="https://example.com/newcerts"; certSha256=*J/lEm9kNRODdCmINbvitpvdYKNQ+YgBj99DlYp4fEXw; date=1511128380; expires=1511733180, srisig; sig=*lGZVaJJM5f2oGczFlLmBdKTDL+QADza4BgeO494ggACYJOvrof6uh5OJCcwKrk7DK+LBch0jssDYPp5CLc1SDA integrity="mi"; validityUrl="https://example.com/resource.validity.1511128380"; ed25519Key=*zsSevyFsxyZHiUluVBDd4eypdRLTqyWRVOJuuKUz+A8 date=1511128380; expires=1511733180, thirdpartysig; sig=*MEYCIQCNxJzn6Rh2fNxsobktir8TkiaJYQFhWTuWI1i4PewQaQIhAMs2TVjc4rTshDtXbgQEOwgj2mRXALhfXPztXgPupii+; integrity="mi"; validityUrl="https://thirdparty.example.com/resource.validity.1511161860"; certUrl="https://thirdparty.example.com/certs"; certSha256=*UeOwUPkvxlGRTyvHcsMUN0A2oNsZbU8EUvg8A9ZAnNc; date=1511133060; expires=1511478660,
There are 4 signatures: 2 from different secp256r1 certificates within https://example.com/, one using a raw ed25519 public key that’s also controlled by example.com, and a fourth using a secp256r1 certificate owned by thirdparty.example.com.
All 4 signatures rely on the MI response header to guard the integrity of the response payload. This isn’t strictly required—some signatures could use MI while others use Digest—but there’s not much benefit to mixing them.
The signatures include a “validityUrl” that includes the first time the resource was seen. This allows multiple versions of a resource at the same URL to be updated with new signatures, which allows clients to avoid transferring extra data while the old versions don’t have known security bugs.
The certificates at https://example.com/oldcerts and https://example.com/newcerts have subjectAltNames of example.com, meaning that if they and their signatures validate, the exchange can be trusted as having an origin of https://example.com/. The author might be using two certificates because their readers have disjoint sets of roots in their trust stores.
The author signed with all three certificates at the same time, so they share a validity range: 7 days starting at 2017-11-19 21:53 UTC.
The author then requested an additional signature from thirdparty.example.com, which did some validation or processing and then signed the resource at 2017-11-19 23:11 UTC. thirdparty.example.com only grants 4-day signatures, so clients will need to re-validate more often.
[I-D.ietf-httpbis-header-structure] provides a way to parameterise labels but not other supported types like binary content. If the Signature header field is notionally a list of parameterised signatures, maybe we should add a “parameterised binary content” type.
Should the certUrl and validityUrl be lists so that intermediates can offer a cache without losing the original URLs? Putting lists in dictionary fields is more complex than [I-D.ietf-httpbis-header-structure] allows, so they’re single items for now.
To sign an exchange’s headers, they need to be serialized into a byte string. Since intermediaries and distributors might rearrange, add, or just reserialize headers, we can’t use the literal bytes of the headers as this serialization. Instead, this section defines a CBOR representation that can be embedded into other CBOR, canonically serialized (Section 3.4), and then signed.
The CBOR representation of an exchange exchange’s headers is the CBOR ([RFC7049]) array with the following content:
Given the HTTP exchange:
GET https://example.com/ HTTP/1.1 Accept: */* HTTP/1.1 200 Content-Type: text/html Digest: SHA-256=20addcf7368837f616d549f035bf6784ea6d4bf4817a3736cd2fc7a763897fe3 Signed-Headers: "content-type", "digest" <!doctype html> <html> ...
The cbor representation consists of the following item, represented using the extended diagnostic notation from [I-D.ietf-cbor-cddl] appendix G:
[
{
':url': 'https://example.com/'
':method': 'GET',
},
{
'digest': 'SHA-256=20addcf7368837f616d549f035bf6784ea6d4bf4817a3736cd2fc7a763897fe3',
':status': '200',
'content-type': 'text/html'
}
]
The resource at a signature’s certUrl MUST have the application/cert-chain+cbor content type, MUST be canonically-encoded CBOR (Section 3.4), and MUST match the following CDDL:
cert-chain = [
"📜⛓", ; U+1F4DC U+26D3
+ {
cert: bytes,
? ocsp: bytes,
? sct: bytes,
* tstr => any,
}
]
The first item in the CBOR array is treated as the end-entity certificate, and the client will attempt to build a path ([RFC5280]) to it from a trusted root using the other certificates in the chain.
Loading a certUrl takes a forceFetch flag. The client MUST:
Within this specification, the canonical serialization of a CBOR item uses the following rules derived from Section 3.9 of [RFC7049] with erratum 4964 applied:
Note: this specification does not use floating point, tags, or other more complex data types, so it doesn’t need rules to canonicalize those.
The client MUST parse the Signature header field as the list of parameterised values (Section 4.8.1 of [I-D.ietf-httpbis-header-structure]) described in Section 3.1. If an error is thrown during this parsing or any of the requirements described there aren’t satisfied, the exchange has no valid signatures. Otherwise, each member of this list represents a signature with parameters.
The client MUST use the following algorithm to determine whether each signature with parameters is invalid or potentially-valid for an exchange. Potentially-valid results include:
This algorithm accepts a forceFetch flag that avoids the cache when fetching URLs. A client that determines that a potentially-valid certificate chain is actually invalid due to an expired OCSP response MAY retry with forceFetch set to retrieve an updated OCSP from the original server.
Set
signing-alg to the result of applying this function to the type of main-certificate’s public key. If the function is undefined on this input, return “invalid”.Note that the above algorithm can determine that an exchange’s headers are potentially-valid before the exchange’s payload is received. Similarly, if integrity identifies a header field like MI ([I-D.thomson-http-mice]) that can incrementally validate the payload, early parts of the payload can be determined to be potentially-valid before later parts of the payload. Higher-level protocols MAY process parts of the exchange that have been determined to be potentially-valid as soon as that determination is made but MUST NOT process parts of the exchange that are not yet potentially-valid. Similarly, as the higher-level protocol determines that parts of the exchange are actually valid, the client MAY process those parts of the exchange and MUST wait to process other parts of the exchange until they too are determined to be valid.
Should the signed message use the TLS format (with an initial 64 spaces) even though these certificates can’t be used in TLS servers?
Both OCSP responses and signatures are designed to expire a short time after they’re signed, so that revoked certificates and signed exchanges with known vulnerabilities are distrusted promptly.
This specification provides no way to update OCSP responses by themselves. Instead, clients need to re-fetch the “certUrl” to get a chain including a newer OCSP response.
The “validityUrl” parameter of the signatures provides a way to fetch new signatures or learn where to fetch a complete updated exchange.
Each version of a signed exchange SHOULD have its own validity URLs, since each version needs different signatures and becomes obsolete at different times.
The resource at a “validityUrl” is “validity data”, a CBOR map matching the following CDDL ([I-D.ietf-cbor-cddl]):
validity = {
? signatures: [ + bytes ]
? update: {
? size: uint,
}
]
The elements of the signatures array are parameterised labels (Section 4.4 of [I-D.ietf-httpbis-header-structure]) meant to replace the signatures within the Signature header field pointing to this validity data. If the signed exchange contains a bug severe enough that clients need to stop using the content, the signatures array MUST NOT be present.
If the the update map is present, that indicates that a new version of the signed exchange is available at its effective request URI (Section 5.5 of [RFC7230]) and can give an estimate of the size of the updated exchange (update.size). If the signed exchange is currently the most recent version, the update SHOULD NOT be present.
If both the signatures and update fields are present, clients can use the estimated size to decide whether to update the whole resource or just its signatures.
For example, say a signed exchange whose URL is https://example.com/resource has the following Signature header field (with line breaks included and irrelevant fields omitted for ease of reading).
Signature: sig1; sig=*MEUCIQ...; ... validityUrl="https://example.com/resource.validity.1511157180"; certUrl="https://example.com/oldcerts"; date=1511128380; expires=1511733180, sig2; sig=*MEQCIG...; ... validityUrl="https://example.com/resource.validity.1511157180"; certUrl="https://example.com/newcerts"; date=1511128380; expires=1511733180, thirdpartysig; sig=*MEYCIQ...; ... validityUrl="https://thirdparty.example.com/resource.validity.1511161860"; certUrl="https://thirdparty.example.com/certs"; date=1511478660; expires=1511824260
At 2017-11-27 11:02 UTC, sig1 and sig2 have expired, but thirdpartysig doesn’t exipire until 23:11 that night, so the client needs to fetch https://example.com/resource.validity.1511157180 (the validityUrl of sig1 and sig2) to update those signatures. This URL might contain:
{
"signatures": [
'sig1; '
'sig=*MEQCIC/I9Q+7BZFP6cSDsWx43pBAL0ujTbON/+7RwKVk+ba5AiB3FSFLZqpzmDJ0NumNwN04pqgJZE99fcK86UjkPbj4jw; '
'validityUrl="https://example.com/resource.validity.1511157180"; '
'integrity="mi"; '
'certUrl="https://example.com/newcerts"; '
'certSha256=*J/lEm9kNRODdCmINbvitpvdYKNQ+YgBj99DlYp4fEXw; '
'date=1511733180; expires=1512337980'
],
"update": {
"size": 5557452
}
}
This indicates that the client could fetch a newer version at https://example.com/resource (the original URL of the exchange), or that the validity period of the old version can be extended by replacing the first two of the original signatures (the ones with a validityUrl of https://example.com/resource.validity.1511157180) with the single new signature provided. (This might happen at the end of a migration to a new root certificate.) The signatures of the updated signed exchange would be:
Signature: sig1; sig=*MEQCIC...; ... validityUrl="https://example.com/resource.validity.1511157180"; certUrl="https://example.com/newcerts"; date=1511733180; expires=1512337980, thirdpartysig; sig=*MEYCIQ...; ... validityUrl="https://thirdparty.example.com/resource.validity.1511161860"; certUrl="https://thirdparty.example.com/certs"; date=1511478660; expires=1511824260
https://example.com/resource.validity.1511157180 could also expand the set of signatures if its signatures array contained more than 2 elements.
Signature header fields cost on the order of 300 bytes for ECDSA signatures, so servers might prefer to avoid sending them to clients that don’t intend to use them. A client can send the Accept-Signature header field to indicate that it does intend to take advantage of any available signatures and to indicate what kinds of signatures it supports.
When a server receives an Accept-Signature header field in a client request, it SHOULD reply with any available Signature header fields for its response that the Accept-Signature header field indicates the client supports. However, if the Accept-Signature value violates a requirement in this section, the server MUST behave as if it hadn’t received any Accept-Signature header at all.
The Accept-Signature header field is a Structured Header as defined by [I-D.ietf-httpbis-header-structure]. Its value MUST be a list (Section 4.8 of [I-D.ietf-httpbis-header-structure]) of parameterised labels (Section 4.4 of [I-D.ietf-httpbis-header-structure]). The order of labels in the Accept-Signature list is not significant. Labels, ignoring any initial “-“ character, MUST NOT be duplicated.
Each label in the Accept-Signature header field’s value indicates that a feature of the Signature header field (Section 3.1) is supported. If the label begins with a “-“ character, it instead indicates that the feature named by the rest of the label is not supported. Unknown labels and parameters MUST be ignored because new labels and new parameters on existing labels may be defined by future specifications.
Labels starting with “digest/” indicate that the client supports the Digest header field ([RFC3230]) with the digest-algorithm from the https://www.iana.org/assignments/http-dig-alg/http-dig-alg.xhtml registry named in lower-case by the rest of the label. For example, “digest/sha-512” indicates support for the SHA-512 digest algorithm, and “-digest/sha-256” indicates non-support for the SHA-256 digest algorithm.
Labels starting with “mi/” indicate that the client supports the MI header field ([I-D.thomson-http-mice]) with the parameter from the HTTP MI Parameter Registry registry named in lower-case by the rest of the label. For example, “mi/mi-blake2” indicates support for Merkle integrity with the as-yet-unspecified mi-blake2 parameter, and “-digest/mi-sha256” indicates non-support for Merkle integrity with the mi-sha256 content encoding.
If the Accept-Signature header field is present, servers SHOULD assume support for “digest/sha-256” and “mi/mi-sha256” unless the header field states otherwise.
Labels starting with “rsa/” indicate that the client supports certificates holding RSA public keys with a number of bits indicated by the digits after the “/”.
Labels starting with “ecdsa/” indicate that the client supports certificates holding ECDSA public keys on the curve named in lower-case by the rest of the label.
If the Accept-Signature header field is present, servers SHOULD assume support for “rsa/2048”, “ecdsa/secp256r1”, and “ecdsa/secp384r1” unless the header field states otherwise.
The “ed25519key” label has parameters indicating the public keys that will be used to validate the returned signature. Each parameter’s name is re-interpreted as binary content (Section 4.5 of [I-D.ietf-httpbis-header-structure]) encoding a prefix of the public key. For example, if the client will validate signatures using the public key whose base64 encoding is 11qYAYKxCrfVS/7TyWQHOg7hcvPapiMlrwIaaPcHURo, valid Accept-Signature header fields include:
Accept-Signature: ..., ed25519key; *11qYAYKxCrfVS/7TyWQHOg7hcvPapiMlrwIaaPcHURo Accept-Signature: ..., ed25519key; *11qYAYKxCrfVS/7TyWQHOg Accept-Signature: ..., ed25519key; *11qYAQ Accept-Signature: ..., ed25519key; *
but not
Accept-Signature: ..., ed25519key; *11qYA
because 5 bytes isn’t a valid length for encoded base64, and not
Accept-Signature: ..., ed25519key; 11qYAQ
because it doesn’t start with the * that indicates binary content.
Note that ed25519key; * is an empty prefix, which matches all public keys, so it’s useful in subresource integrity (Appendix A.3) cases like <link rel=preload as=script href="..."> where the public key isn’t known until the matching <script src="..." integrity="..."> tag.
Accept-Signature: mi/mi-sha256
states that the client will accept signatures with payload integrity assured by the MI header and mi-sha256 content encoding and implies that the client will accept integrity assured by the Digest: SHA-256 header and signatures from 2048-bit RSA keys and ECDSA keys on the secp256r1 and secp384r1 curves.
Accept-Signature: -rsa/2048, rsa/4096
states that the client will accept 4096-bit RSA keys but not 2048-bit RSA keys, and implies that the client will accept ECDSA keys on the secp256r1 and secp384r1 curves and payload integrity assured with the MI: mi-sha256 and Digest: SHA-256 header fields.
Is an Accept-Signature header useful enough to pay for itself? If clients wind up sending it on most requests, that may cost more than the cost of sending Signatures unconditionally. On the other hand, it gives servers an indication of which kinds of signatures are supported, which can help us upgrade the ecosystem in the future.
Is Accept-Signature the right spelling, or do we want to imitate Want-Digest (Section 4.3.1 of [RFC3230]) instead?
Do I have the right structure for the labels indicating feature support?
To determine whether to trust a cross-origin exchange, the client takes a Signature header field (Section 3.1) and the exchange. The client MUST parse the Signature header into a list of signatures according to the instructions in Section 3.5, and run the following algorithm for each signature, stopping at the first one that returns “valid”. If any signature returns “valid”, return “valid”. Otherwise, return “invalid”.
As described in Section 6.1, a publisher can cause problems if they sign an exchange that includes private information. There’s no way for a client to be sure an exchange does or does not include private information, but header fields that store or convey stored state in the client are a good sign.
A stateful request header field informs the server of per-client state. These include but are not limited to:
A stateful response header field modifies state, including authentication status, in the client. The HTTP cache is not considered part of this state. These include but are not limited to:
We define a new X.509 extension, CanSignHttpExchanges to be used in the certificate when the certificate permits the usage of signed exchanges. When this extension is not present the client MUST NOT accept a signature from the certificate as proof that a signed exchange is authoritative for a domain covered by the certificate. When it is present, the client MUST follow the validation procedure in Section 4.
id-ce-canSignHttpExchanges OBJECT IDENTIFIER ::= { TBD }
CanSignHttpExchanges ::= BIT STRING { allowed (0) }
Leaf certificates without this extension need to be revoked if the private key is exposed to an unauthorized entity, but they generally don’t need to be revoked if a signing oracle is exposed and then removed.
CA certificates, by contrast, need to be revoked if an unauthorized entity is able to make even one unauthorized signature.
Certificates with this extension MUST be revoked if an unauthorized entity is able to make even one unauthorized signature.
Conforming CAs MUST mark this extension as critical, and clients MUST NOT accept certificates with this extension in TLS connections (Section 4.4.2.2 of [I-D.ietf-tls-tls13]). This prevents accidental signing oracles exposed by TLS servers from allowing package signing (e.g. [DROWN] and [ROBOT]).
A signed exchange can be transferred in several ways, of which three are described here.
The signature for a signed exchange can be included in a normal HTTP response. Because different clients send different request header fields, and intermediate servers add response header fields, it can be impossible to have a signature for the exact request and response that the client sees. Therefore, when a client validates the Signature header field for an exchange represented as a normal HTTP request/response pair, it MUST pass only the subset of header fields defined by Section 5.1.1 to the validation procedure (Section 3.5).
If the client relies on signature validity for any aspect of its behavior, it MUST ignore any header fields that it didn’t pass to the validation procedure.
The significant headers of an exchange represented as a normal HTTP request/response pair (Section 2.1 of [RFC7230] or Section 8.1 of [RFC7540]) are:
If the exchange’s Signed-Headers header field is not present, doesn’t parse as a Structured Header ([I-D.ietf-httpbis-header-structure]) or doesn’t follow the constraints on its value described in Section 5.1.2, the exchange has no significant headers.
Do the significant headers of an exchange need to include the Signed-Headers header field itself?
The Signed-Headers header field identifies an ordered list of response header fields to include in a signature. The request URL and response status are included unconditionally. This allows a TLS-terminating intermediate to reorder headers without breaking the signature. This can also allow the intermediate to add headers that will be ignored by some higher-level protocols, but Section 3.5 provides a hook to let other higher-level protocols reject such insecure headers.
This header field appears once instead of being incorporated into the signatures’ parameters because the signed header fields need to be consistent across all signatures of an exchange, to avoid forcing higher-level protocols to merge the header field lists of valid signatures.
See Appendix B.2 for a discussion of why only the URL from the request is included and not other request headers.
Signed-Headers is a Structured Header as defined by [I-D.ietf-httpbis-header-structure]. Its value MUST be a list (Section 4.8 of [I-D.ietf-httpbis-header-structure]) of lowercase strings (Section 4.2 of [I-D.ietf-httpbis-header-structure]) naming HTTP response header fields. Pseudo-header field names (Section 8.1.2.1 of [RFC7540]) MUST NOT appear in this list.
Higher-level protocols SHOULD place requirements on the minimum set of headers to include in the Signed-Headers header field.
To allow servers to Server-Push (Section 8.2 of [RFC7540]) signed exchanges (Section 3) signed by an authority for which the server is not authoritative (Section 9.1 of [RFC7230]), this section defines an HTTP/2 extension.
Clients that might accept signed Server Pushes with an authority for which the server is not authoritative indicate this using the HTTP/2 SETTINGS parameter ENABLE_CROSS_ORIGIN_PUSH (0xSETTING-TBD).
An ENABLE_CROSS_ORIGIN_PUSH value of 0 indicates that the client does not support cross-origin Push. A value of 1 indicates that the client does support cross-origin Push.
A client MUST NOT send a ENABLE_CROSS_ORIGIN_PUSH setting with a value other than 0 or 1 or a value of 0 after previously sending a value of 1. If a server receives a value that violates these rules, it MUST treat it as a connection error (Section 5.4.1 of [RFC7540]) of type PROTOCOL_ERROR.
The use of a SETTINGS parameter to opt-in to an otherwise incompatible protocol change is a use of “Extending HTTP/2” defined by Section 5.5 of [RFC7540]. If a server were to send a cross-origin Push without first receiving a ENABLE_CROSS_ORIGIN_PUSH setting with the value of 1 it would be a protocol violation.
The signatures on a Pushed cross-origin exchange may be untrusted for several reasons, for example that the certificate could not be fetched, that the certificate does not chain to a trusted root, that the signature itself doesn’t validate, that the signature is expired, etc. This draft conflates all of these possible failures into one error code, NO_TRUSTED_EXCHANGE_SIGNATURE (0xERROR-TBD).
How fine-grained should this specification’s error codes be?
If the client has set the ENABLE_CROSS_ORIGIN_PUSH setting to 1, the server MAY Push a signed exchange for which it is not authoritative, and the client MUST NOT treat a PUSH_PROMISE for which the server is not authoritative as a stream error (Section 5.4.2 of [RFC7540]) of type PROTOCOL_ERROR, as described in Section 8.2 of [RFC7540].
Instead, the client MUST validate such a PUSH_PROMISE and its response by taking the Signature header field from the response, and the exchange consisting of the PUSH_PROMISE and the response without that Signature header field, and passing them to the algorithm in Section 4. If this returns “invalid”, the client MUST treat the response as a stream error (Section 5.4.2 of [RFC7540]) of type NO_TRUSTED_EXCHANGE_SIGNATURE. Otherwise, the client MUST treat the pushed response as if the server were authoritative for the PUSH_PROMISE’s authority.
Is it right that “validityUrl” is required to be same-origin with the exchange? This allows the mitigation against downgrades in Section 6.3, but prohibits intermediates from providing a cache of the validity information. We could do both with a list of URLs.
To allow servers to serve cross-origin responses when either the client or the server hasn’t implemented HTTP/2 Push (Section 8.2 of [RFC7540]) support yet, we define a format that represents an HTTP exchange.
The application/http-exchange+cbor content type encodes an HTTP exchange, including request metadata and header fields, optionally a request body, response header fields and metadata, a payload body, and optionally trailer header fields.
This content type consists of a canonically-serialized (Section 3.4) CBOR array containing:
This specification defines the following member names with their associated values:
A parser MAY return incremental information while parsing application/http-exchange+cbor content.
Members “request”, “response”, and “payload” MUST be present. If one is missing, the parser MUST stop and report an error.
The member names MUST appear in the order:
If a member name is not a text string, appears out of order, or is followed by a value not matching its description above, the parser MUST stop and report an error.
If the parser encounters an unknown member name, it MUST skip the following item and resume parsing at the next member name.
An example application/http-exchange+cbor file representing a possible exchange with https://example.com/ follows, in the extended diagnostic format defined in Appendix G of [I-D.ietf-cbor-cddl]:
[
"htxg",
"request",
{
':method': 'GET',
':url': 'https://example.com/',
'accept', '*/*'
},
"response",
{
':status': '200',
'content-type': 'text/html'
},
"payload",
'<!doctype html>\r\n<html>...'
]
Should application/http-exchange+cbor support request payloads and trailers, or only the aspects needed for signed exchanges?
Are the mime type, extension, and magic number right?
If a publisher blindly signs all responses as their origin, they can cause at least two kinds of problems, described below. To avoid this, publishers SHOULD design their systems to opt particular public content that doesn’t depend on authentication status into signatures instead of signing by default.
Signing systems SHOULD also incorporate the following mitigations to reduce the risk that private responses are signed:
Blind signing can sign responses that create session cookies or otherwise change state on the client to identify a particular session. This breaks certain kinds of CSRF defense and can allow an attacker to force a user into the attacker’s account, where the user might unintentionally save private information, like credit card numbers or addresses.
This specification defends against cookie-based attacks by blocking the Set-Cookie response header, but it cannot prevent Javascript or other response content from changing state.
If a site signs private information, an attacker might set up their own account to show particular private information, forward that signed information to a victim, and use that victim’s confusion in a more sophisticated attack.
Stripping authentication information from requests before sending them to backends is likely to prevent the backend from showing attacker-specific information in the signed response. It does not prevent the attacker from showing their victim a signed-out page when the victim is actually signed in, but while this is still misleading, it seems less likely to be useful to the attacker.
Relaxing the requirement to consult DNS when determining authority for an origin means that an attacker who possesses a valid certificate no longer needs to be on-path to redirect traffic to them; instead of modifying DNS, they need only convince the user to visit another Web site in order to serve responses signed as the target. This consideration and mitigations for it are shared by the combination of [I-D.ietf-httpbis-origin-frame] and [I-D.ietf-httpbis-http2-secondary-certs].
Signing a bad response can affect more users than simply serving a bad response, since a served response will only affect users who make a request while the bad version is live, while an attacker can forward a signed response until its signature expires. Authors should consider shorter signature expiration times than they use for cache expiration times.
Clients MAY also check the “validityUrl” of an exchange more often than the signature’s expiration would require. Doing so for an exchange with an HTTPS request URI provides a TLS guarantee that the exchange isn’t out of date (as long as Section 5.2.3.1 is resolved to keep the same-origin requirement).
An attacker with temporary access to a signing oracle can sign “still valid” assertions with arbitrary timestamps and expiration times. As a result, when a signing oracle is removed, the keys it provided access to MUST be revoked so that, even if the attacker used them to sign future-dated exchange validity assertions, the key’s OCSP assertion will expire, causing the exchange as a whole to become untrusted.
The use of a single Signed-Headers header field prevents us from signing aspects of the request other than its effective request URI (Section 5.5 of [RFC7230]). For example, if an author signs both Content-Encoding: br and Content-Encoding: gzip variants of a response, what’s the impact if an attacker serves the brotli one for a request with Accept-Encoding: gzip?
The simple form of Signed-Headers also prevents us from signing less than the full request URL. The SRI use case (Appendix A.3) may benefit from being able to leave the authority less constrained.
Section 3.5 can succeed when some delivered headers aren’t included in the signed set. This accommodates current TLS-terminating intermediates and may be useful for SRI (Appendix A.3), but is risky for trusting cross-origin responses (Appendix A.1, Appendix A.2, and Appendix A.6). Section 5.2 requires all headers to be included in the signature before trusting cross-origin pushed resources, at Ryan Sleevi’s recommendation.
Clients MUST NOT trust an effective request URI claimed by an application/http-exchange+cbor resource (Section 5.3) without either ensuring the resource was transferred from a server that was authoritative (Section 9.1 of [RFC7230]) for that URI’s origin, or passing the Signature response header field from the exchange stored in the resource, and that exchange without its Signature response header field, to the procedure in Section 4, and getting “valid” back.
Normally, when a client fetches https://o1.com/resource.js, o1.com learns that the client is interested in the resource. If o1.com signs resource.js, o2.com serves it as https://o2.com/o1resource.js, and the client fetches it from there, then o2.com learns that the client is interested, and if the client executes the Javascript, that could also report the client’s interest back to o1.com.
Often, o2.com already knew about the client’s interest, because it’s the entity that directed the client to o1resource.js, but there may be cases where this leaks extra information.
For non-executable resource types, a signed response can improve the privacy situation by hiding the client’s interest from the original author.
To prevent network operators other than o1.com or o2.com from learning which exchanges were read, clients SHOULD only load exchanges fetched over a transport that’s protected from eavesdroppers. This can be difficult to determine when the exchange is being loaded from local disk, but when the client itself requested the exchange over a network it SHOULD require TLS ([I-D.ietf-tls-tls13]) or a successor transport layer, and MUST NOT accept exchanges transferred over plain HTTP without TLS.
TODO: possibly register the validityUrl format.
This section registers the Signature header field in the “Permanent Message Header Field Names” registry ([RFC3864]).
Header field name: Signature
Applicable protocol: http
Status: standard
Author/Change controller: IETF
Specification document(s): Section 3.1 of this document
This section establishes an entry for the HTTP/2 Settings Registry that was established by Section 11.3 of [RFC7540]
Name: ENABLE_CROSS_ORIGIN_PUSH
Code: 0xSETTING-TBD
Initial Value: 0
Specification: This document
This section establishes an entry for the HTTP/2 Error Code Registry that was established by Section 11.4 of [RFC7540]
Name: NO_TRUSTED_EXCHANGE_SIGNATURE
Code: 0xERROR-TBD
Description: The client does not trust the signature for a cross-origin Pushed signed exchange.
Specification: This document
Type name: application
Subtype name: http-exchange+cbor
Required parameters: N/A
Optional parameters: N/A
Encoding considerations: binary
Security considerations: see Section 6.6
Interoperability considerations: N/A
Published specification: This specification (see Section 5.3).
Applications that use this media type: N/A
Fragment identifier considerations: N/A
Additional information:
Deprecated alias names for this type: N/A
Magic number(s): 8? 64 68 74 78 67
File extension(s): .htxg
Macintosh file type code(s): N/A
Person and email address to contact for further information: See Authors’ Addresses section.
Intended usage: COMMON
Restrictions on usage: N/A
Author: See Authors’ Addresses section.
Change controller: IESG
Type name: application
Subtype name: cert-chain+cbor
Required parameters: N/A
Optional parameters: N/A
Encoding considerations: binary
Security considerations: N/A
Interoperability considerations: N/A
Published specification: This specification (see Section 3.3).
Applications that use this media type: N/A
Fragment identifier considerations: N/A
Additional information:
Deprecated alias names for this type: N/A
Magic number(s): 1*9(??) 67 F0 9F 93 9C E2 9B 93
File extension(s): N/A
Macintosh file type code(s): N/A
Person and email address to contact for further information: See Authors’ Addresses section.
Intended usage: COMMON
Restrictions on usage: N/A
Author: See Authors’ Addresses section.
Change controller: IESG
| [DROWN] | Aviram, N., Schinzel, S., Somorovsky, J., Heninger, N., Dankel, M., Steube, J., Valenta, L., Adrian, D., Halderman, J., Dukhovni, V., Käsper, E., Cohney, S., Engels, S., Paar, C. and Y. Shavitt, "The DROWN Attack", 2016. |
| [I-D.burke-content-signature] | Burke, B., "HTTP Header for digital signatures", Internet-Draft draft-burke-content-signature-00, March 2011. |
| [I-D.cavage-http-signatures] | Cavage, M. and M. Sporny, "Signing HTTP Messages", Internet-Draft draft-cavage-http-signatures-09, November 2017. |
| [I-D.ietf-httpbis-http2-secondary-certs] | Bishop, M., Sullivan, N. and M. Thomson, "Secondary Certificate Authentication in HTTP/2", Internet-Draft draft-ietf-httpbis-http2-secondary-certs-00, December 2017. |
| [I-D.ietf-httpbis-origin-frame] | Nottingham, M. and E. Nygren, "The ORIGIN HTTP/2 Frame", Internet-Draft draft-ietf-httpbis-origin-frame-06, January 2018. |
| [I-D.thomson-http-content-signature] | Thomson, M., "Content-Signature Header Field for HTTP", Internet-Draft draft-thomson-http-content-signature-00, July 2015. |
| [I-D.yasskin-webpackage-use-cases] | Yasskin, J., "Use Cases and Requirements for Web Packages", Internet-Draft draft-yasskin-webpackage-use-cases-00, August 2017. |
| [RFC2965] | Kristol, D. and L. Montulli, "HTTP State Management Mechanism", RFC 2965, DOI 10.17487/RFC2965, October 2000. |
| [RFC6066] | Eastlake 3rd, D., "Transport Layer Security (TLS) Extensions: Extension Definitions", RFC 6066, DOI 10.17487/RFC6066, January 2011. |
| [RFC6265] | Barth, A., "HTTP State Management Mechanism", RFC 6265, DOI 10.17487/RFC6265, April 2011. |
| [RFC6454] | Barth, A., "The Web Origin Concept", RFC 6454, DOI 10.17487/RFC6454, December 2011. |
| [RFC6455] | Fette, I. and A. Melnikov, "The WebSocket Protocol", RFC 6455, DOI 10.17487/RFC6455, December 2011. |
| [RFC7235] | Fielding, R. and J. Reschke, "Hypertext Transfer Protocol (HTTP/1.1): Authentication", RFC 7235, DOI 10.17487/RFC7235, June 2014. |
| [RFC7615] | Reschke, J., "HTTP Authentication-Info and Proxy-Authentication-Info Response Header Fields", RFC 7615, DOI 10.17487/RFC7615, September 2015. |
| [RFC8053] | Oiwa, Y., Watanabe, H., Takagi, H., Maeda, K., Hayashi, T. and Y. Ioku, "HTTP Authentication Extensions for Interactive Clients", RFC 8053, DOI 10.17487/RFC8053, January 2017. |
| [ROBOT] | Böck, H., Somorovsky, J. and C. Young, "The ROBOT Attack", 2017. |
| [SRI] | Akhawe, D., Braun, F., Marier, F. and J. Weinberger, "Subresource Integrity", World Wide Web Consortium Recommendation REC-SRI-20160623, June 2016. |
| [W3C.NOTE-OPS-OverHTTP] | Hensley, P., Metral, M., Shardanand, U., Converse, D. and M. Myers, "Implementation of OPS Over HTTP", W3C NOTE NOTE-OPS-OverHTTP, June 1997. |
To reduce round trips, a server might use HTTP/2 Push (Section 8.2 of [RFC7540]) to inject a subresource from another server into the client’s cache. If anything about the subresource is expired or can’t be verified, the client would fetch it from the original server.
For example, if https://example.com/index.html includes
<script src="https://jquery.com/jquery-1.2.3.min.js">
Then to avoid the need to look up and connect to jquery.com in the critical path, example.com might push that resource signed by jquery.com.
In order to speed up loading but still maintain control over its content, an HTML page in a particular origin O.com could tell clients to load its subresources from an intermediate content distributor that’s not authoritative, but require that those resources be signed by O.com so that the distributor couldn’t modify the resources. This is more constrained than the common CDN case where O.com has a CNAME granting the CDN the right to serve arbitrary content as O.com.
<img logicalsrc="https://O.com/img.png"
physicalsrc="https://distributor.com/O.com/img.png">
To make it easier to configure the right distributor for a given request, computation of the physicalsrc could be encapsulated in a custom element:
<dist-img src="https://O.com/img.png"></dist-img>
where the <dist-img> implementation generates an appropriate <img> based on, for example, a <meta name="dist-base"> tag elsewhere in the page. However, this has the downside that the preloader can no longer see the physical source to download it. The resulting delay might cancel out the benefit of using a distributor.
This could be used for some of the same purposes as SRI (Appendix A.3).
To implement this with the current proposal, the distributor would respond to the physical request to https://distributor.com/O.com/img.png with first a signed PUSH_PROMISE for https://O.com/img.png and then a redirect to https://O.com/img.png.
The W3C WebAppSec group is investigating using signatures in [SRI]. They need a way to transmit the signature with the response, which this proposal provides.
Their needs are simpler than most other use cases in that the integrity="ed25519-[public-key]" attribute and CSP-based ways of expressing a public key don’t need that key to be wrapped into a certificate.
The “ed25519Key” signature parameter supports this simpler way of attaching a key.
The current proposal for signature-based SRI describes signing only the content of a resource, while this specification requires them to sign the request URI as well. This issue is tracked in https://github.com/mikewest/signature-based-sri/issues/5. The details of what they need to sign will affect whether and how they can use this proposal.
So-called “Binary Transparency” may eventually allow users to verify that a program they’ve been delivered is one that’s available to the public, and not a specially-built version intended to attack just them. Binary transparency systems don’t exist yet, but they’re likely to work similarly to the successful Certificate Transparency logs described by [RFC6962].
Certificate Transparency depends on Signed Certificate Timestamps that prove a log contained a particular certificate at a particular time. To build the same thing for Binary Transparency logs containing HTTP resources or full websites, we’ll need a way to provide signatures of those resources, which signed exchanges provides.
Native app stores like the Apple App Store and the Android Play Store grant their contents powerful abilities, which they attempt to make safe by analyzing the applications before offering them to people. The web has no equivalent way for people to wait to run an update of a web application until a trusted authority has vouched for it.
While full application analysis probably needs to wait until the authority can sign bundles of exchanges, authorities may be able to guarantee certain properties by just checking a top-level resource and its [SRI]-constrained sub-resources.
Fully-offline websites can be represented as bundles of signed exchanges, although an optimization to reduce the number of signature verifications may be needed. Work on this is in progress in the https://github.com/WICG/webpackage repository.
To verify that a thing came from a particular origin, for use in the same context as a TLS connection, we need someone to vouch for the signing key with as much verification as the signing keys used in TLS. The obvious way to do this is to re-use the web PKI and CA ecosystem.
If we re-use existing TLS server certificates, we incur the risks that:
These risks are considered too high, so we define a new X.509 certificate extension in Section 4.2 that requires CAs to issue new certificates for this purpose. We expect at least one low-cost CA to be willing to sign certificates with this extension.
In order to prevent an attacker who can convince the server to sign some resource from causing those signed bytes to be interpreted as something else the new X.509 extension here is forbidden from being used in TLS servers. If Section 4.2 changes to allow re-use in TLS servers, we would need to:
The specification also needs to define which signing algorithm to use. It currently specifies that as a function from the key type, instead of allowing attacker-controlled data to specify it.
The client needs to be able to find the certificate vouching for the signing key, a chain from that certificate to a trusted root, and possibly other trust information like SCTs ([RFC6962]). One approach would be to include the certificate and its chain in the signature metadata itself, but this wastes bytes when the same certificate is used for multiple HTTP responses. If we decide to put the signature in an HTTP header, certificates are also unusually large for that context.
Another option is to pass a URL that the client can fetch to retrieve the certificate and chain. To avoid extra round trips in fetching that URL, it could be bundled with the signed content or PUSHed with it. The risks from the client_certificate_url extension (Section 11.3 of [RFC6066]) don’t seem to apply here, since an attacker who can get a client to load an exchange and fetch the certificates it references, can also get the client to perform those fetches by loading other HTML.
To avoid using an unintended certificate with the same public key as the intended one, the content of the leaf certificate or the chain should be included in the signed data, like TLS does (Section 4.4.3 of [I-D.ietf-tls-tls13]).
The previous [I-D.thomson-http-content-signature] and [I-D.burke-content-signature] schemes signed just the content, while ([I-D.cavage-http-signatures] could also sign the response headers and the request method and path. However, the same path, response headers, and content may mean something very different when retrieved from a different server. Section 5.1.1 currently includes the whole request URL in the signature, but it’s possible we need a more flexible scheme to allow some higher-level protocols to accept a less-signed URL.
The question of whether to include other request headers—primarily the accept* family—is still open. These headers need to be represented so that clients wanting a different language, say, can avoid using the wrong-language response, but it’s not obvious that there’s a security vulnerability if an attacker can spoof them. For now, the proposal (Section 3) omits other request headers.
In order to allow multiple clients to consume the same signed exchange, the exchange shouldn’t include the exact request headers that any particular client sends. For example, a Japanese resource wouldn’t include
accept-language: ja-JP, ja;q=0.9, en;q=0.8, zh;q=0.7, *;q=0.5
Instead, it would probably include just
accept-language: ja-JP, ja
and clients would use the same matching logic as for PUSH_PROMISE frame headers.
HTTP headers are traditionally munged by proxies, making it impossible to guarantee that the client will see the same sequence of bytes as the author wrote. In the HTTPS world, we have more end-to-end header integrity, but it’s still likely that there are enough TLS-terminating proxies that the author’s signatures would tend to break before getting to the client.
There’s also no way in current HTTP for the response to a client-initiated request (Section 8.1 of [RFC7540]) to convey the request headers it expected to respond to. A PUSH_PROMISE (Section 8.2 of [RFC7540]) does not have this problem, and it would be possible to introduce a response header to convey the expected request headers.
Since proxies are unlikely to modify unknown content types, we can wrap the original exchange into an application/http-exchange+cbor format (Section 5.3) and include the Cache-Control: no-transform header when sending it.
To reduce the likelihood of accidental modification by proxies, the application/http-exchange+cbor format includes a file signature that doesn’t collide with other known signatures.
To help the PUSHed subresources use case (Appendix A.1), we might also want to extend the PUSH_PROMISE frame type to include a signature, and that could tell intermediates not to change the ensuing headers.
A normal HTTPS response is authoritative only for one client, for as long as its cache headers say it should live. A signed exchange can be re-used for many clients, and if it was generated while a server was compromised, it can continue compromising clients even if their requests happen after the server recovers. This signing scheme needs to mitigate that risk.
Certificates are mis-issued and private keys are stolen, and in response clients need to be able to stop trusting these certificates as promptly as possible. Online revocation checks don’t work, so the industry has moved to pushed revocation lists and stapled OCSP responses [RFC6066].
Pushed revocation lists work as-is to block trust in the certificate signing an exchange, but the signatures need an explicit strategy to staple OCSP responses. One option is to extend the certificate download (Appendix B.1.3) to include the OCSP response too, perhaps in the TLS 1.3 CertificateEntry format.
The signed content in a response might be vulnerable to attacks, such as XSS, or might simply be discovered to be incorrect after publication. Once the author fixes those vulnerabilities or mistakes, clients should stop trusting the old signed content in a reasonable amount of time. Similar to certificate revocation, I expect the best option to be stapled “this version is still valid” assertions with short expiration times.
These assertions could be structured as:
The signature also needs to include instructions to intermediates for how to fetch updated validity assertions.
This draft could expire signature validity using the normal HTTP cache control headers ([RFC7234]) instead of embedding an expiration date in the signature itself. This section specifies how that would work, and describes why I haven’t chosen that option.
The signatures in the Signature header field (Section 3.1) would no longer contain “date” or “expires” fields.
The validity-checking algorithm (Section 3.5) would initialize date from the resource’s Date header field (Section 7.1.1.2 of [RFC7231]) and initialize expires from either the Expires header field (Section 5.3 of [RFC7234]) or the Cache-Control header field’s max-age directive (Section 5.2.2.8 of [RFC7234]) (added to date), whichever is present, preferring max-age (or failing) if both are present.
Validity updates (Section 3.6) would include a list of replacement response header fields. For each header field name in this list, the client would remove matching header fields from the stored exchange’s response header fields. Then the client would append the replacement header fields to the stored exchange’s response header fields.
For example, given a stored exchange of:
GET https://example.com/ HTTP/1.1 Accept: */* HTTP/1.1 200 Date: Mon, 20 Nov 2017 10:00:00 UTC Content-Type: text/html Date: Tue, 21 Nov 2017 10:00:00 UTC Expires: Sun, 26 Nov 2017 10:00:00 UTC <!doctype html> <html> ...
And an update listing the following headers:
Expires: Fri, 1 Dec 2017 10:00:00 UTC Date: Sat, 25 Nov 2017 10:00:00 UTC
The resulting stored exchange would be:
GET https://example.com/ HTTP/1.1 Accept: */* HTTP/1.1 200 Content-Type: text/html Expires: Fri, 1 Dec 2017 10:00:00 UTC Date: Sat, 25 Nov 2017 10:00:00 UTC <!doctype html> <html> ...
In an exchange with multiple signatures, using cache control to expire signatures forces all signatures to initially live for the same period. Worse, the update from one signature’s “validityUrl” might not match the update for another signature. Clients would need to maintain a current set of headers for each signature, and then decide which set to use when actually parsing the resource itself.
This need to store and reconcile multiple sets of headers for a single signed exchange argues for embedding a signature’s lifetime into the signature.
RFC EDITOR PLEASE DELETE THIS SECTION.
draft-03
draft-02
Thanks to Ilari Liusvaara, Justin Schuh, Mark Nottingham, Mike Bishop, Ryan Sleevi, and Yoav Weiss for comments that improved this draft.