I recently had to fix a Server-Side Request Forgery bug in Libravatar's OpenID support. In addition to enabling authentication on internal services whenever possible, I also forced all outgoing network requests from the Django web-application to go through a restrictive egress proxy.

OpenID logins are prone to SSRF

Server-Side Request Forgeries are vulnerabilities which allow attackers to issue arbitrary GET requests on the server side. Unlike a Cross-Site Request Forgery, SSRF requests do not include user credentials (e.g. cookies). On the other hand, since these requests are done by the server, they typically originate from inside the firewall.

This allows attackers to target internal resources and issue arbitrary GET requests to them. One could use this to leak information, especially when error reports include the request payload, tamper with the state of internal services or portscan an internal network.

OpenID 1.x logins are prone to these vulnerabilities because of the way they are initiated:

Users visit a site's login page.
They enter their OpenID URL in a text field.
The server fetches the given URL to discover the OpenID endpoints.
The server redirects the user to their OpenID provider to continue the rest of the login flow.

The third step is the potentially problematic one since it requires a server-side fetch.

Filtering URLs in the application is not enough

At first, I thought I would filter out undesirable URLs inside the application:

hostnames like localhost, 127.0.0.1 or ::1
non-HTTP schemes like file or gopher
non-standard ports like 5432 or 11211

However this filtering is going to be very easy to bypass:

Add a hostname in your DNS zone which resolves to 127.0.0.1.
Setup a redirect to a blacklisted URL such as file:///etc/passwd.

Applying the filter on the original URL is clearly not enough.

Install and configure a Squid proxy

In order to fully restrict outgoing OpenID requests from the web application, I used a Squid HTTP proxy.

First, install the package:

apt install squid3

and set the following in /etc/squid3/squid.conf:

acl to_localnet dst 0.0.0.1-0.255.255.255 # RFC 1122 "this" network (LAN)
acl to_localnet dst 10.0.0.0/8            # RFC 1918 local private network (LAN)
acl to_localnet dst 100.64.0.0/10         # RFC 6598 shared address space (CGN)
acl to_localnet dst 169.254.0.0/16        # RFC 3927 link-local (directly plugged) machines
acl to_localnet dst 172.16.0.0/12         # RFC 1918 local private network (LAN)
acl to_localnet dst 192.168.0.0/16        # RFC 1918 local private network (LAN)
acl to_localnet dst fc00::/7              # RFC 4193 local private network range
acl to_localnet dst fe80::/10             # RFC 4291 link-local (directly plugged) machines

acl SSL_ports port 443
acl Safe_ports port 80
acl Safe_ports port 443
acl CONNECT method CONNECT

http_access deny !Safe_ports
http_access deny CONNECT !SSL_ports
http_access deny manager
http_access deny to_localhost
http_access deny to_localnet
http_access allow localhost
http_access deny all

http_port 127.0.0.1:3128

cache deny all

Ideally, I would like to use a whitelist approach to restrict requests to a small set of valid URLs, but in the case of OpenID, the set of valid URLs is not fixed. Therefore the only workable approach is a blacklist. The above snippet whitelists port numbers (80 and 443) and blacklists requests to localhost (a built-in squid acl variable which resolves to 127.0.0.1 and ::1) as well as known local IP ranges.

Expose the proxy to Django in the WSGI configuration

In order to force all outgoing requests from Django to go through the proxy, I put the following in my WSGI application (/etc/libravatar/django.wsgi):

os.environ['ftp_proxy'] = "http://127.0.0.1:3128"
os.environ['http_proxy'] = "http://127.0.0.1:3128"
os.environ['https_proxy'] = "http://127.0.0.1:3128"

The whole thing seemed to work well in my limited testing. There is however a bug in urllib2 with proxying HTTPS URLs that include a port number, and there is an open issue in python-openid around proxies and OpenID.