Convert the initial Google Doc to Markdown.

pull/2454/head
Patrick Cloke 4 years ago
parent 7666b9fea7
commit c69a747bae

@ -0,0 +1,262 @@
# User-Interactive Auth for SSO-backed homeservers
## Background
Certain endpoints, such as `DELETE /_matrix/client/r0/devices/{deviceId}` and
`POST /_matrix/client/r0/account/3pid/add`, require the user to reconfirm their
identity, as a guard against a leaked access token being used to take over an
entire account.
On a normal homeserver, this is done by prompting the user to enter their
password. However, on a homeserver where users authenticate via a single-sign-on
system, the user doesn't have a password, so this doesn't work. Instead we need
to delegate that check to the SSO system.
At the protocol level, this means adding support for SSO to
[user-interactive auth](https://matrix.org/docs/spec/client_server/r0.6.0#user-interactive-authentication-api).
### The current implementation
Or, "how UI Auth works, in practice":
When the client calls one of the protected endpoints, it initially returns a 401
response. For example:
```
POST /_matrix/client/r0/delete_devices HTTP/1.1
Content-Type: application/json
{
"devices": ["FSVVTZRRAA"]
}
HTTP/1.1 401 Unauthorized
Content-Type: application/json
{
"flows": [
{
"stages": [
"m.login.password"
]
}
],
"params": {},
"session": "dTKfsLHSAJeAhqfxUsvrIVJd"
}
```
The client:
* inspects the "flows" list
* discovers there is a flow it knows how to follow
* carries out the first "stage" of that flow (m.login.password)
ie, the client prompts the user to enter a password.
The client then resubmits with an additional 'auth' param, with "type" giving
the name of the authentication type it has just carried out. That completes the
authentication flow, so the server is now happy, and returns a 200:
```
POST /_matrix/client/r0/delete_devices HTTP/1.1
Content-Type: application/json
{
"devices": ["FSVVTZRRAA"],
"auth": {
"session":"dTKfsLHSAJeAhqfxUsvrIVJd",
"type":"m.login.password",
"identifier":{"type":"m.id.user", "user":"@userid:matrix.org"},
"password":"<password>"
}
}
HTTP/1.1 200 OK
Content-Type: application/json
Content-Length: 179
{}
```
## Proposal
We add an `m.login.sso` authentication type to the UI auth spec. There are no
additional params as part of this auth type.
1. If the client wants to complete that authentication type, it opens a browser
window for `/_matrix/client/r0/auth/m.login.sso/fallback/web?session=<...>`
with session set to the UI-Auth session id (from the "auth" dict).
The homeserver returns a page which says words to the effect of "A client is
trying to remove a device/add an email address/take over your account. To
confirm this action, **re-authenticate with single sign-on**. If you did not
expect this, your account may be compromised!"
See security section below.
2. The link, once the user clicks on it, goes to the SSO provider's page.
3. The SSO provider validates the user, and redirects the browser back to the
homeserver.
4. The homeserver validates the response from the SSO provider, updates the
user-interactive auth session to show that the SSO has completed, serves the
fallback auth completion page as specced:
https://matrix.org/docs/spec/client_server/r0.6.0#fallback
5. The client resubmits its original request, with its original session id,
which now hopefully completes.
Note that the post-SSO URL on the homeserver is left up to the homeserver
implementation rather than forming part of the spec.
* SAML2 servers typically only support one URL per service provider, so in
practice it will need to be the same as that already used for the login flow
(for synapse, it's /_matrix/saml2/authn_response) - and the server needs to
be able to figure out if it's doing SSO for a login attempt or an SSO
attempt.
* CAS doesn't have the same restriction.
### Example flow:
0. Client submits the request, which the server says requires SSO:
```
POST /_matrix/client/r0/delete_devices HTTP/1.1
Content-Type: application/json
Authorization: Bearer xyzzy
{
"devices": ["FSVVTZRRAA"]
}
HTTP/1.1 401 Unauthorized
Content-Type: application/json
{
"flows": [
{
"stages": [
"m.login.sso"
]
}
],
"params": {},
"session": "dTKfsLHSAJeAhqfxUsvrIVJd"
}
```
1. Client opens a browser window for the fallback endpoint:
```
GET /_matrix/client/r0/auth/m.login.sso/fallback/web
?session=dTKfsLHSAJeAhqfxUsvrIVJd HTTP/1.1
HTTP/1.1 200 OK
<body>
can delete device pls?
<a href="https://sso_provider/validate?SAMLRequest=...">clicky</a>
</body>
```
2. The user clicks the confirmation link which goes to the SSO provider's site:
```
GET https://sso_provider/validate?SAMLRequest=<etc> HTTP/1.1
... SAML/CAS stuff
```
3. The SSO provider validates the user and ends up redirecting the browser back
to the homeserver. (The example below shows a 302 for simplicity but SAML normally uses a 200 with an html <form>, with javascript to automatically submit it.)
```
HTTP/1.1 302 Moved
Location: https://homeserver/_matrix/saml2/authn_response?
SAMLResponse=<etc>
```
4. The browser sends the SSO response to the homeserver, which validates it and
shows the fallback auth completion page:
```
GET /_matrix/saml2/authn_response?SAMLResponse=<etc>
HTTP/1.1 200 OK
<script>
if (window.onAuthDone) {
window.onAuthDone();
} else if (window.opener && window.opener.postMessage) {
window.opener.postMessage("authDone", "*");
}
</script>
<p>Thank you.</p>
<p>You may now close this window and return to the application.</p>
```
5. The client closes the browser popup if it is still open, and resubmits its
original request, which now succeeds:
```
POST /_matrix/client/r0/delete_devices HTTP/1.1
Content-Type: application/json
Authorization: Bearer xyzzy
{
"auth": {
"session": "dTKfsLHSAJeAhqfxUsvrIVJd"
}
}
HTTP/1.1 200 OK
Content-Type: application/json
{}
```
A few notes:
* The security of this relies on UI-Auth sessions only being used for the same
request as they were initiated for. I don't think that's currently enforced.
* We might consider an alternative client flow where the fallback auth ends up
redirecting to a given URI, instead of doing javascript postMessage foo. I
think that's an orthogonal change to the fallback auth though.
* In theory, any clients that already implement the fallback process for
unknown authentication types will work fine without modification.
Unfortunately, I don't think Riot (on any platform) is among them.
## Security considerations
Recall that the thing we are trying to guard against here is a single leaked
access-token being used to take over an entire account. So let's assume the
attacker has got hold of an access token for your account. What happens if we
skip the confirmation step?
The attacker, who has your access token, starts a UI-Auth session to add his
email address to your account.
He then sends you a link "hey, check out this cool video!"; the link leads (via
an innocent-looking url shortener) to
`/_matrix/client/r0/auth/m.login.sso/fallback/web?session=<...>`, with the ID of
the session that he just created.
Recall that we're skipping the confirmation step, so the server redirects
straight to the SSO provider.
It's common for SSO providers to redirect straight back to the app if you've
recently authenticated with them; even in the best case, the SSO provider shows
some innocent message along the lines of "Confirm that you want to sign in to
<your matrix homeserver>".
So the SSO completes, and the attacker's session is validated.
By telling the user what's about to happen as clearly as we can, and making them
confirm, we can mitigate this problem.
## Unstable prefix
We should use a vendor prefix here until this hits the spec.
`org.matrix.login.sso` ?
Loading…
Cancel
Save