COMMAND
sockd
SYSTEMS AFFECTED
Those using socks
PROBLEM
Gerhard Rieger found following. In connections established via
the SOCKS 4 protocol the client chooses the address of the target
application server. Therefore NEC recommends to configure sockd
to deny any connections to the internal network if used in
conjunction with a firewall. But what if the client requests a
connection to the loopback (127.0.0.1) address? It appears that
the only protection of the socks server's loopback interface is
on the CLIENT side! The socks client software connects 127.0.0.1
directly to the client's machine loopback interface. If someone
changes the appropriate lines in the socks client code, then the
client sends the request - strictly as directed by its
/etc/socks.conf - to the socks server. sockd might permit this
connection, if it has a rather liberal /etc/sockd.conf
configuration. It will then connect to the requested TCP port on
the socks server host's loopback interface.
This allows for three typical exploit scenarios:
1) The client can circumvent IP filters that protect the
firewall's services on the network interfaces.
2) The client can reach TCP services that are listening only
on the loopback interface.
3) The client can circumvent IP address based authentication,
because the accessed service only sees the loopback address
with which sockd connected instead of the real client's
address.
4) Scenario where the (socks) proxy itself is listening on all
interfaces, loopback as well as external. Connect to
proxy, authenticate with proxy, socks_connect to localhost,
authenticate, socks_connect, wash rinse repeat. In most
cases this will result in some form of resource exhaustion,
or as in some cases with Wingate's most recent version of
SOCKS, server death and even OS death (hang or blue
screen). Also, this isn't just limited to loopback; you're
actually more likely to accomplish this form of DoS by
reconnecting to the proxy's external IP. And, even if you
block these kinds of "looped" connections, you can easily
bounce between two SOCKS servers to accomplish the same
attack. Of course, for any of the above to be an easily
perpetrated attack, the SOCKS server has to allow AUTH_NONE
(method 0x00) authentication. Unfortunately, by default
most SOCKS servers do (although, IIRC Dante is not one of
these).
In an experiment to connect from a UNIX client on the internal
network to port 6000 on a typically configured SOCKS 4 based UNIX
firewall with IP filters and the X server running. The X server
would not serve connection request from client machine due to
strict host access based authentication settings and IP filter
protection. But when used X Windows over a manipulated socks
client to connect to the socks server on the firewall and having
it access port 6000 on 127.0.0.1, tester succeed to take, e.g., a
screendump from the firewall display.
SOLUTION
Everybody who gets to know about this problem can simply protect
his socks server by adding a line to the beginning of
/etc/sockd.conf:
deny 0.0.0.0 0.0.0.0 127.0.0.0 255.0.0.0
The reference implementation of socks5 handles the situation
properly. More importantly, SOCKS users must configure their ACLs
properly. Having a line like "permit - - - - - - -" is asking for
security problems.