Summary

Using OAuth 2 to authenticate with Synapse has many benefits over using older authentication mechanisms. We can try to consolidate authentication services to using OAuth flows wherever possible. The next step is to permit the creation of OAuth 2 public clients, which would allow OAuth 2 based authentication in Synapse command line apps and browser-based SPAs like the GWT web client.

To securely support OAuth 2 public clients, these actions are needed (backfill this list with Jira tickets when confirmed):

Support PKCE in the authorization code flow
Add field to label OAuth 2 clients as either “public” or “confidential”; backfill existing clients as “confidential”
- Public clients should not be issued secrets
- Public clients should be required to use PKCE
Revoke the current iteration of a refresh token when an old refresh token is used

Background

There are currently many ways to authenticate a request to Synapse, including

Session tokens. Session tokens grant access to all account functions. Session tokens expire after 24 hours, and they can also be refreshed (for an additional 24 hours) and revoked. To acquire a session token, users must either
- Enter their username and password
- Sign in using Google as an OIDC identity provider.
API keys. API keys grant access to all account functions. API keys do not expire, but they can be revoked, and a new key can be generated. Users may only have one API key active at a time. A user may retrieve their API key at any time (and as an implication, they are stored unhashed in Synapse).
Using an access token. Access tokens last 24 hours. Access tokens are also scoped, so that a token may only be used to perform specific types of actions. Only verified OAuth 2 clients may be issued access tokens. At this time, access tokens cannot be revoked unless they are issued with a refresh token.
- There are two ways that a client can obtain an access token, both requiring client credentials.
  - Sending Synapse a valid authorization code, which is granted when a user authorizes the OAuth client. Authorization codes expire after one minute.
  - Sending Synapse a valid refresh token. The refresh token is granted when the user authorizes an app with the offline_access scope. Refresh tokens expire 180 days after being issued and are single-use, though a new refresh token is issued when one is used, so access should be considered non-expiring. Refresh tokens can be revoked, and access tokens associated with/granted by the refresh token will also be revoked. Additionally, we only store the hash of the refresh token, so the token can only be retrieved at the time it is generated.

Table Summary

	Session Token	API Key	OAuth 2.0 access/refresh token
Access Granted	Entire account	Entire account	Scoped, defined by the client
Expires	24 hrs after being issued	Forever	Access: 24 hrs after issued Refresh: 180 days after issued
Revocable	Yes	Yes	Only if associated with refresh token
Maximum that can be issued	1	1	Access: Unlimited Refresh: 100 per user per OAuth client
Stored as/Re-retrievable	Unhashed/Yes	Unhashed/Yes	Access: Not stored/No Refresh: Hashed token/No

Motivation

We can reduce the complexity and attack surface of authentication with Synapse by utilizing OAuth 2 login flows in places where other mechanisms are currently used.

In particular, we can aim to replace usage of the API key with OAuth refresh tokens. The most common use case for the API key is to authenticate Synapse command line clients. Using refresh/access tokens in place of an API key provides a couple of advantages:

Leased/timed expiration - if a particular session isn’t used for 180 days, the token expires
The ability to grant/revoke access on the machine-level, instead of having one key used everywhere
Scoped access, in cases where the command line app is running a job that only needs a subset of access/functionality
Using either the authorization code (currently supported) or device code (not currently supported) OAuth 2 flows require users to login on Synapse.org via a browser, instead of entering credentials at the command line, or pasting an API key into a config file.

While currently a lower priority, we can also consider the ability to replace the usage of session tokens with access tokens in the GWT client as the backend requirements are very similar.

The caveat to using OAuth 2.0 for native apps and JavaScript-based SPAs is that in these circumstances, the OAuth client cannot securely and confidentially store credentials. This scenario is a use case that is outlined in the OAuth 2.0 specification, but we currently expect all Synapse OAuth clients to be confidential clients. Therefore, we must consider the security and access implications of creating/allowing public OAuth clients.

Requirements/Notes/Considerations

Each point in this section will be fleshed out with reasoning, pros, cons, etc. and then consolidated once confident in what’s necessary.

PKCE

Best practices suggest that authorization servers are required to support PKCE [2][3]. PKCE only provides benefits to public OAuth clients, since a malicious app must have knowledge of an OAuth client’s credentials to execute the attack.

PKCE in a nutshell

PKCE prevents a malicious application from hijacking an authorization code. This can occur when using a native app’s OAuth client, and the redirect URI is compromised by a malicious application observing traffic to the user agent (mobile app or web browser). In this scenario, the malicious app has the authorization code, and is in possession of the client’s credentials (if they exist), because it is a public client. Thus the malicious app now has access to the user’s account.

With PKCE, the client generates a high-entropy random string, and calculates the SHA256 hash of the string. When making the initial authorization request, the hash is sent to the authorization server. The authorization server associates the hash with the authorization code, before sending it to the client (the hash may be embedded)

When a client uses the authorization code, they must provide the initial random string. The authorization server computes the SHA256 hash of the string, and will only issue token(s) to the client if the hashes match.

Designating new clients as either public or confidential

All existing clients are considered to be “confidential” clients, so we can easily backfill that information.

We can then use the public/confidential label to handle logic accordingly [2]:

Authorization servers MUST record the client type in the client registration details in order to identify and process requests accordingly.

As an example, we can require that public clients utilize PKCE, among other considerations.

In OAuth 2.0, the authorization server SHOULD NOT assume client type [1]. We can require client creators to specify their client type upon creation. This is a breaking API change, but acceptable because OAuth clients aren’t created programmatically (it wouldn’t make sense if they are, since they must be manually verified to be functional).

Public clients may or may not be issued a secret

Client credentials (other than the ID) are not useful for public clients [1]:

The authorization server MAY establish a client authentication method with public clients. However, the authorization server MUST NOT rely on public client authentication for the purpose of identifying the client.

OAuth 2.1 ([2]) suggests that they shouldn’t be required, so why issue them?

Secrets that are statically included as part of an app distributed to multiple users should not be treated as confidential secrets, as one user may inspect their copy and learn the shared secret. For this reason, it is NOT RECOMMENDED for authorization servers to require client authentication of public native apps clients using a shared secret, as this serves little value beyond client identification which is already provided by the "client_id" request parameter.

The only endpoints that require OAuth client credentials involve issuing or revoking tokens, so permitting public clients to have no secret should not cause any issues with, for example, public client metadata being editable by unauthorized parties.

Public clients may be required to use certain grant types

Currently, the only grant types implemented are the more secure possible grant types, likely meaning no restrictions at this time.
Users may need to explicitly opt-in to allowing public client access to their resources (e.g. a simple checkbox and supplemental info modal on their profile settings page)

Notes on public clients using refresh tokens

OAuth 2.1 ([2]) Note on refresh tokens for public clients:

Authorization server MUST utilize one of these methods to detect refresh token replay by malicious actors for public clients:
* _Sender-constrained refresh tokens:_ the authorization server cryptographically binds the refresh token to a certain client instance by utilizing [I-D.ietf-oauth-token-binding] or [RFC8705]
* _Refresh token rotation:_ the authorization server issues a new refresh token with every access token refresh response. The previous refresh token is invalidated but information about the relationship is retained by the authorization server. If a refresh token is compromised and subsequently used by both the attacker and the legitimate client, one of them will present an invalidated refresh token, which will inform the authorization server of the breach. The authorization server cannot determine which party submitted the invalid refresh token, but it will revoke the active refresh token. This stops the attack at the cost of forcing the legitimate client to obtain a fresh authorization grant.

Token binding introduces a public/private keypair and encrypts additional data to bind tokens to a TLS connection, thus a token could not be used if intercepted via MitM/token export. While powerful, implementing token binding can be complicated and has seen low adoption. We may choose to further investigate token binding as in [4] later. Token binding as described in [OAuth 2.0 Token Binding] does not seem to be an acceptable solution, because the draft has expired, and the mechanism to implement token binding on the client side does not seem to be supported by any major browsers (abandoned by Chrome, and only implemented in Edge before it became Chromium-based).
Refresh token rotation is implemented, but we currently do not revoke an active token if an invalid token is used. We should extend our implementation to include this behavior.