Slashdot Mirror
Overconfidence in SSH Protection
nitsudima writes to mention a post on the Informit site about the common misunderstandings surrounding SSH, and how well-intentioned admins may be creating holes in their own security by using it. From the article: "In UNIX, all things are files. To send network traffic, UNIX writes the traffic to the network device file. In this case, the connection to Box A (and that private key used for authentication) is a socket file. This file will shuttle the authentication traffic between Box A and Box P. So what's the risk? Maybe the hacker can't get a copy of the private key through the socket file, but something better (from his/her view) can be done. If the hacker has root on Box D, he or she can point a private copy of the agent forwarding software to that socket file and thereby point the authentication process to the administrator's credentials--the ones kept on the 'safe' intranet. What are the chances that the administrator has configured access to all the DMZ servers he controls?"
The submitter didn't summarise anything, he cut out a chunk which didn't make much sense on its own. It didn't help that the article was fairly long-winded. This is what I understand the author is trying to say:
As others have commented, this is kind of a "duh" moment. What's the next article?Gan Family Homepage
The article illustrated one very convoluted way to break your DMZ security, but failed to make the simple statement: don't trust anyone, not even root, on your DMZ hosts. Allow SSH logins into the DMZ, and allow the DMZ to pull files from private network patching servers, such as apt repositories, but don't allow anyone to SSH from the DMZ to the intranet. Assume the DMZ is cracked wide open and keystroke logging. No one is going to get past the DMZ by watching you type 'apt-get install squid' but they will by watching you type 'ssh root@creditcarddb.int' and then the root password.
... you can packet sniff the authentication process all day long and you won't get someone's private key.
Anyone who tunnels from the DMZ to a trusted host which can execute commands on a sensitive server can't see the forest for the trees. You've learned how to use SSH and tunnel, but you're lacking some basic common sense.
Also, I don't see what good a socket catching the authentication will do
That whole article seemed a bit of voodoo itself. Many incongruous statements, like "If the hacker has root on Box D, he or she can point a private copy of the agent forwarding software to that socket file and thereby point the authentication process to the administrator's credentials--the ones kept on the "safe" intranet."
What does that mean, exactly? You direct the authentication process to a socket file and point the process to the admin's credentials? If the socket is on the DMZ host, and the credentials are on the private network host, how can you point the authentication process to those credentials?
Maybe I'm stupid, but the article didn't seem to make a lot of sense.
Here's a summary for you:
User A on box foo:
foo> ssh-agent xterm
foo> ssh-add
* enters their pass key *
User A can now ssh to any box that has their public key in box:$HOME/.ssh/authorized_keys
User B (evul hacker with root on box foo):
foo# SSH_AGENT_PID=XXXX; export SSH_AGENT_PID
foo# SSH_AUTH_SOCK=/tmp/ssh-YYYY/ZZZZ; export SSH_AUTH_SOCK
User B now can ssh to any box that User A can, as above.
(where XXXX, YYYY, and ZZZZ are determined by evul hacker)
It's not just you. I had to re-read it several times.
I think the main point (the one the article submitter picked up on) was that if an attacker can compromise your DMZ box (the most vulnerable box your company owns and hence the least trusted box your company owns) that has no private ssh keys stored on it and can't connect to any other trusted box but does have trusted boxes connecting to it, then he can use that to compromise further trusted boxes inside the organisation.
To put it another way, if you ssh to an attacker's machine using agent forwarding he can probably ssh back to yours.
Sig matters not. Judge me by my sig, do you?
YES Thats correct you can use AgentForwarding.... If you are stupid enough to use agent forwarding to a host you don't trust or you would consider insecure ITS YOUR OWN STUPID FAULT IF YOU GET HACKED. Now for the evil h4x0rz to use agent forwarding on the host you connect to to hack the machine you are coming in from requires quite a number of things to be done on your stupid behalf that sure wouldnt be enabled by default and you would almost need to set them up purposefully. The only real danger with agent forwarding to an insucure host is that evil h4x0rz on that host can use your forwarded authentication agent to connect to boxes that are set up to both allow connections using that ssh-key AND allow tcp connections from any box that the evil h4x0rz have access to. Aside from that it is only as insecure as establishing a telnet session to the box and having some buffer overflow occur back to the client due to poor code on the client side. I am sure not about to stop using ssh for some "simpler" protocol like telnet but I will sure keep disabling AgentForwarding and any kind of portforwarding the hosts I dont trust and I ASSUME EVERYONE ELSE WILL CONTINUE TO DO THAT AS WELL. Otherwise you might as well start posting your root passwords to slashdot which may or may not matter if you have locked your systems down correctly in the first place.
meridian at tha.net
Why the developers of ssh have an option to forbid agent forwarding. Isn't it off by default? I cite from "man ssh":
>>>
Agent forwarding should be enabled with caution. Users with the
ability to bypass file permissions on the remote host (for the
agent's Unix-domain socket) can access the local agent through
the forwarded connection. An attacker cannot obtain key material
from the agent, however they can perform operations on the keys
that enable them to authenticate using the identities loaded into
the agent.
So wha is slashdot running an article about something where there is an explicit one-paragraph long waring in the man page of program at the option in question.
Yes, no doutbt there are a lot of idiots around, who without understanding,do things which require semantics which leads to a security leak (there is abolutely no way if you want to initiate authenticatication from processes on a machine to avoid root to do the same - as log as you are not asked on the agent's side each time before authentication;
1) Don't allow the DMZ box to ssh anywhere; firewall it off. There should be no need to ssh FROM the DMZ box, only TO it.
Or better yet, don't allow the DMZ host to initiate *any* connection outbound from itself, if the services present on it don't need to do such, or failing that, disallow it initiating any connection that isn't out the internet-only-facing interface(s).
However, that's still not what the attack is exploiting, and wouldn't prevent the attack.
The 'attack' is taking an (I)Internal (S)erver, an (I)nternal (W)orkstation, and a (D)MZ host. Your firewall/ACLs are set such that IS can't receive, or will reject, any connection from D. However IS will accept connections from IW, and furthermore IW can ssh to D. So, the well-meaning admin of D, using IW, ssh's to D, setting up a tunnel to forward traffic on port Dx on D back to IW, port IWx, and also ssh's from IW to IS, forwarding IWx to a service on IS. Thus you can now connection to (D's) localhost:Dx and end up talking to the service on IS.
At no point is any connection initiated from D outside of itself, as the data is simply passing back through the ssh tunnel from IW to D, and then back further from IW to IS. And, no, you can't firewall D from talking to any but necessary ports on D, as we're assuming root compromise of D and thus all such bets are off.
Now, if someone has compromised D *and* can hijack this tunnel D IW IS they have access to IS.
Of course the real solution to the base problem is to have IS set up in some way to push data out to D, such that IW's user/D's admin doesn't have to play such silly and dangerous games in the first place. Any such 'administrator' setting up what has been described is incompetent.
Now one last thing. The general attack hinges on an attacker's agent Aa being able to make use of the unix domain socket of the administrator's agent, Da. I'm very certain that when I tried this kind of attack on myself way back in 1998 or sooner it plain wasn't possible. If it is now then the (Open, whatever)ssh code has taken a step backwards. Basically some check was done on the origin of the messages on the socket, and if they weren't as expected the request to use the keys in the agent was denied. I think it was along the lines of "is the requesting process a (sub(sub...))child of myself?", presumably by following parent process IDs back up until it finds itself or init. Yes, ok, if the agent spawned any child that spawned a service that was subsequently compromised and not put in a new session group you could probably pull off the attack, but that is unlikely (as any service daemonising itself will end up in a new session group).
ssh-agent is a solution to grant any process on the system full access to a means of authenticating through your private key.
Any process? Presumably this is only if you have deliberately made the socket files world-readable?
The alternatives to ssh-agent are password authentication and host-based authentication. I'm sure I don't need to remind you about the security issues with those!
The article is about a common misconfiguration with regard to agent forwarding. The DMZ hosts aren't supposed to be safe, that's why they're in the DMZ and not in the intranet. The admin must assume that root on these machines is compromised. Consequently he doesn't store his private keys on any of the DMZ machines. But what many overlook, possibly because they don't use the feature, is agent forwarding. Once the admin has logged into a compromised DMZ host, access to his credentials is extended to the DMZ host by that ominous socket. The file itself never leaves the admin's computer, but if agent forwarding is enabled, root on the DMZ host can now point other hosts on the intranet to the authentication facility on the admin's computer. This misconfiguration enables the attacker to hop from the DMZ to the intranet. The correct way to avoid this is to disable agent forwarding (on the admin's computer, not just on the DMZ hosts, of course).
Perhaps the author of the article should have read the source of the text you quoted. The preceding paragraph:
So the only people who will be caught out by this are those who:
Configuring the agent to prompt the user to confirm any signing request can be as complicated as putting the private key on a smart card (which will make the reader prompt for a PIN whenever the card recieves a signing request) or it can be as simple as using the -c option when calling ssh-add; therefore this does not seem like a big deal to me.
A few versions ago OpenSSH added a -c "Require confirmation to sign using identities" to ssh-add to take care of this. Or using something like SSHKeychain on OS X so it'll ask for confirmation for multi-hop auth, but not for connections direct from your trusted machine.
Tom Anthony
No, it's not really a man in the middle attack.
It's more of a credential hijacking scenario where
the attacker waits for you to authenticate with
the compromised machine, forward your credentials
to that machine, and then the attacker uses those
credentials to reach other machines that honor those
credentials.
This would be more like you signing in, walking
away from your computer, and someone else walkup up
to the computer and doing stuff as you except that
they get to act as you while you're still acting
as you.
Did that help?
*sigh* back to work...
Don't use SSH to poke a hole in the firewall separating your DMZ from the intranet.
DMZ hosts are typically configured so that they can NOT connect out to hosts outside the DMZ. This is enforced by firewall restrictions. As long as this is done, there is no problem.
A slashdotter who didn't build his own computer is like a Jedi who didn't build his own lightsaber.
Care to explain how a remote root login is any more dangerous than a remote user login when the user can sudo to root?
A slashdotter who didn't build his own computer is like a Jedi who didn't build his own lightsaber.
-A Enables forwarding of the authentication agent connection. This can also be specified on a per-host basis in a configuration file.
Agent forwarding should be enabled with caution. Users with the ability to bypass file permissions on the remote host (for the agent's Unix-domain socket) can access the local agent through the forwarded connection. An attacker cannot obtain key material from the agent, however they can perform operations on the keys that enable them to authenticate using the identities loaded into the agent.
So what's the news?