matrix-spec

You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

18 KiB

Raw Blame History

title	weight	type
Identity Service API	40	docs

The Matrix client-server and server-server APIs are largely expressed in Matrix user identifiers. From time to time, it is useful to refer to users by other ("third-party") identifiers, or "3PID"s, e.g. their email address or phone number. This Identity Service Specification describes how mappings between third-party identifiers and Matrix user identifiers can be established, validated, and used. This description technically may apply to any 3PID, but in practice has only been applied specifically to email addresses and phone numbers.

General principles

The purpose of an identity server is to validate, store, and answer questions about the identities of users. In particular, it stores associations of the form "identifier X represents the same user as identifier Y", where identities may exist on different systems (such as email addresses, phone numbers, Matrix user IDs, etc).

The identity server has some private-public keypairs. When asked about an association, it will sign details of the association with its private key. Clients may validate the assertions about associations by verifying the signature with the public key of the identity server.

In general, identity servers are treated as reliable oracles. They do not necessarily provide evidence that they have validated associations, but claim to have done so. Establishing the trustworthiness of an individual identity server is left as an exercise for the client.

3PID types are described in 3PID Types Appendix.

API standards

The mandatory baseline for identity server communication in Matrix is exchanging JSON objects over HTTP APIs. HTTPS is required for communication, and all API calls use a Content-Type of application/json. In addition, strings MUST be encoded as UTF-8.

Any errors which occur at the Matrix API level MUST return a "standard error response". This is a JSON object which looks like:

{
  "errcode": "<error code>",
  "error": "<error message>"
}

The error string will be a human-readable error message, usually a sentence explaining what went wrong. The errcode string will be a unique string which can be used to handle an error message e.g. M_FORBIDDEN. There may be additional keys depending on the error, but the keys error and errcode MUST always be present.

Some standard error codes are below:

M_NOT_FOUND
The resource requested could not be located.

M_MISSING_PARAMS
The request was missing one or more parameters.

M_INVALID_PARAM
The request contained one or more invalid parameters.

M_SESSION_NOT_VALIDATED
The session has not been validated.

M_NO_VALID_SESSION
A session could not be located for the given parameters.

M_SESSION_EXPIRED
The session has expired and must be renewed.

M_INVALID_EMAIL
The email address provided was not valid.

M_EMAIL_SEND_ERROR
There was an error sending an email. Typically seen when attempting to verify ownership of a given email address.

M_INVALID_ADDRESS
The provided third party address was not valid.

M_SEND_ERROR
There was an error sending a notification. Typically seen when attempting to verify ownership of a given third party address.

M_UNRECOGNIZED
The request contained an unrecognised value, such as an unknown token or medium.

M_THREEPID_IN_USE
The third party identifier is already in use by another user. Typically this error will have an additional mxid property to indicate who owns the third party identifier.

M_UNKNOWN
An unknown error has occurred.

Privacy

Identity is a privacy-sensitive issue. While the identity server exists to provide identity information, access should be restricted to avoid leaking potentially sensitive data. In particular, being able to construct large-scale connections between identities should be avoided. To this end, in general APIs should allow a 3PID to be mapped to a Matrix user identity, but not in the other direction (i.e. one should not be able to get all 3PIDs associated with a Matrix user ID, or get all 3PIDs associated with a 3PID).

Version 1 API deprecation

As described on each of the version 1 endpoints, the v1 API is deprecated in favour of the v2 API described here. The major difference, with the exception of a few isolated cases, is that the v2 API requires authentication to ensure the user has given permission for the identity server to operate on their data.

The v1 API is planned to be removed from the specification in a future version.

Clients SHOULD attempt the v2 endpoints first, and if they receive a 404, 400, or similar error they should try the v1 endpoint or fail the operation. Clients are strongly encouraged to warn the user of the risks in using the v1 API, if they are planning on using it.

Web browser clients

It is realistic to expect that some clients will be written to be run within a web browser or similar environment. In these cases, the identity server should respond to pre-flight requests and supply Cross-Origin Resource Sharing (CORS) headers on all requests.

When a client approaches the server with a pre-flight (OPTIONS) request, the server should respond with the CORS headers for that route. The recommended CORS headers to be returned by servers on all requests are:

Access-Control-Allow-Origin: *
Access-Control-Allow-Methods: GET, POST, PUT, DELETE, OPTIONS
Access-Control-Allow-Headers: Origin, X-Requested-With, Content-Type, Accept, Authorization

Authentication

Most v2 endpoints in the Identity Service API require authentication in order to ensure that the requesting user has accepted all relevant policies and is otherwise permitted to make the request. The v1 API (currently deprecated) does not require this authentication, however using v1 is strongly discouraged as it will be removed in a future release.

Identity Servers use a scheme similar to the Client-Server API's concept of access tokens to authenticate users. The access tokens provided by an Identity Server cannot be used to authenticate Client-Server API requests.

An access token is provided to an endpoint in one of two ways:

Via a query string parameter, access_token=TheTokenHere.
Via a request header, Authorization: Bearer TheTokenHere.

Clients are encouraged to the use the Authorization header where possible to prevent the access token being leaked in access/HTTP logs. The query string should only be used in cases where the Authorization header is inaccessible for the client.

When credentials are required but missing or invalid, the HTTP call will return with a status of 401 and the error code M_UNAUTHORIZED.

Terms of service

Identity Servers are encouraged to have terms of service (or similar policies) to ensure that users have agreed to their data being processed by the server. To facilitate this, an identity server can respond to almost any authenticated API endpoint with an HTTP 403 and the error code M_TERMS_NOT_SIGNED. The error code is used to indicate that the user must accept new terms of service before being able to continue.

All endpoints which support authentication can return the M_TERMS_NOT_SIGNED error. When clients receive the error, they are expected to make a call to GET /terms to find out what terms the server offers. The client compares this to the m.accepted_terms account data for the user (described later) and presents the user with option to accept the still-missing terms of service. After the user has made their selection, if applicable, the client sends a request to POST /terms to indicate the user's acceptance. The server cannot expect that the client will send acceptance for all pending terms, and the client should not expect that the server will not respond with another M_TERMS_NOT_SIGNED on their next request. The terms the user has just accepted are appended to m.accepted_terms.

Status check

Key management

An identity server has some long-term public-private keypairs. These are named in a scheme algorithm:identifier, e.g. ed25519:0. When signing an association, the standard Signing JSON algorithm applies.

The identity server may also keep track of some short-term public-private keypairs, which may have different usage and lifetime characteristics than the service's long-term keys.

Association lookup

Client behaviour

{{% boxes/note %}} This section only covers the v2 lookup endpoint. The v1 endpoint is described in isolation above. {{% /boxes/note %}}

Prior to performing a lookup clients SHOULD make a request to the /hash_details endpoint to determine what algorithms the server supports (described in more detail below). The client then uses this information to form a /lookup request and receive known bindings from the server.

Clients MUST support at least the sha256 algorithm.

Server behaviour

{{% boxes/note %}} This section only covers the v2 lookup endpoint. The v1 endpoint is described in isolation above. {{% /boxes/note %}}

Servers, upon receipt of a /lookup request, will compare the query against known bindings it has, hashing the identifiers it knows about as needed to verify exact matches to the request.

Servers MUST support at least the sha256 algorithm.

Algorithms

Some algorithms are defined as part of the specification, however other formats can be negotiated between the client and server using /hash_details.

`sha256`

This algorithm MUST be supported by clients and servers at a minimum. It is additionally the preferred algorithm for lookups.

When using this algorithm, the client converts the query first into strings separated by spaces in the format <address> <medium> <pepper>. The <pepper> is retrieved from /hash_details, the <medium> is typically email or msisdn (both lowercase), and the <address> is the 3PID to search for. For example, if the client wanted to know about alice@example.org's bindings, it would first format the query as alice@example.org email ThePepperGoesHere.

{{% boxes/rationale %}} Mediums and peppers are appended to the address to prevent a common prefix for each 3PID, helping prevent attackers from pre-computing the internal state of the hash function. {{% /boxes/rationale %}}

After formatting each query, the string is run through SHA-256 as defined by RFC 4634. The resulting bytes are then encoded using URL-Safe Unpadded Base64 (similar to room version 4's event ID format).

An example set of queries when using the pepper matrixrocks would be:

"alice@example.com email matrixrocks" -> "4kenr7N9drpCJ4AfalmlGQVsOn3o2RHjkADUpXJWZUc"
"bob@example.com email matrixrocks"   -> "LJwSazmv46n0hlMlsb_iYxI0_HXEqy_yj6Jm636cdT8"
"18005552067 msisdn matrixrocks"      -> "nlo35_T5fzSGZzJApqu8lgIudJvmOQtDaHtr-I4rU7I"

The set of hashes is then given as the addresses array in /lookup. Note that the pepper used MUST be supplied as pepper in the /lookup request.

`none`

This algorithm performs plaintext lookups on the identity server. Typically this algorithm should not be used due to the security concerns of unhashed identifiers, however some scenarios (such as LDAP-backed identity servers) prevent the use of hashed identifiers. Identity servers (and optionally clients) can use this algorithm to perform those kinds of lookups.

Similar to the sha256 algorithm, the client converts the queries into strings separated by spaces in the format <address> <medium> - note the lack of <pepper>. For example, if the client wanted to know about alice@example.org's bindings, it would format the query as alice@example.org email.

The formatted strings are then given as the addresses in /lookup. Note that the pepper is still required, and must be provided to ensure the client has made an appropriate request to /hash_details first.

Security considerations

{{% boxes/note %}} MSC2134 has much more information about the security considerations made for this section of the specification. This section covers the high-level details for why the specification is the way it is. {{% /boxes/note %}}

Typically the lookup endpoint is used when a client has an unknown 3PID it wants to find a Matrix User ID for. Clients normally do this kind of lookup when inviting new users to a room or searching a user's address book to find any Matrix users they may not have discovered yet. Rogue or malicious identity servers could harvest this unknown information and do nefarious things with it if it were sent in plain text. In order to protect the privacy of users who might not have a Matrix identifier bound to their 3PID addresses, the specification attempts to make it difficult to harvest 3PIDs.

{{% boxes/rationale %}} Hashing identifiers, while not perfect, helps make the effort required to harvest identifiers significantly higher. Phone numbers in particular are still difficult to protect with hashing, however hashing is objectively better than not.

An alternative to hashing would be using bcrypt or similar with many rounds, however by nature of needing to serve mobile clients and clients on limited hardware the solution needs be kept relatively lightweight. {{% /boxes/rationale %}}

Clients should be cautious of servers not rotating their pepper very often, and potentially of servers which use a weak pepper - these servers may be attempting to brute force the identifiers or use rainbow tables to mine the addresses. Similarly, clients which support the none algorithm should consider at least warning the user of the risks in sending identifiers in plain text to the identity server.

Addresses are still potentially reversable using a calculated rainbow table given some identifiers, such as phone numbers, common email address domains, and leaked addresses are easily calculated. For example, phone numbers can have roughly 12 digits to them, making them an easier target for attack than email addresses.

Establishing associations

The flow for creating an association is session-based.

Within a session, one may prove that one has ownership of a 3PID. Once this has been established, the user can form an association between that 3PID and a Matrix user ID. Note that this association is only proved one way; a user can associate any Matrix user ID with a validated 3PID, i.e. I can claim that any email address I own is associated with @billg:microsoft.com.

Sessions are time-limited; a session is considered to have been modified when it was created, and then when a validation is performed within it. A session can only be checked for validation, and validation can only be performed within a session, within a 24-hour period since its most recent modification. Any attempts to perform these actions after the expiry will be rejected, and a new session should be created and used instead.

To start a session, the client makes a request to the appropriate /requestToken endpoint. The identity server then sends a validation token to the user, and the user provides the token to the client. The client then provides the token to the appropriate /submitToken endpoint, completing the session. At this point, the client should /bind the third party identifier or leave it for another entity to bind.

Format of a validation token

The format of the validation token is left up to the identity server: it should choose one appropriate to the 3PID type. (For example, it would be inappropriate to expect a user to copy a long passphrase including punctuation from an SMS message into a client.)

Whatever format the identity server uses, the validation token must consist of at most 255 Unicode codepoints. Clients must pass the token through without modification.

Email associations

Phone number associations

General

Invitation storage

An identity server can store pending invitations to a user's 3PID, which will be retrieved and can be either notified on or look up when the 3PID is associated with a Matrix user ID.

At a later point, if the owner of that particular 3PID binds it with a Matrix user ID, the identity server will attempt to make an HTTP POST to the Matrix user's homeserver via the /3pid/onbind endpoint. The request MUST be signed with a long-term private key for the identity server.

Ephemeral invitation signing

To aid clients who may not be able to perform crypto themselves, the identity server offers some crypto functionality to help in accepting invitations. This is less secure than the client doing it itself, but may be useful where this isn't possible.

18 KiB Raw Blame History