From mboxrd@z Thu Jan 1 00:00:00 1970 From: Steve Grubb Subject: Re: (no subject) Date: Fri, 20 Apr 2007 19:32:21 -0400 Message-ID: <200704201932.21306.sgrubb@redhat.com> References: Mime-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: Content-Disposition: inline List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Sender: linux-audit-bounces@redhat.com Errors-To: linux-audit-bounces@redhat.com To: linux-audit@redhat.com List-Id: linux-audit@redhat.com On Friday 20 April 2007 18:13:17 paul moore wrote: > My understanding is that the auid/loginid process property is to allow the > audit system to *really* know who did things In particular it seems to be > for tracking who did things when they run su or sudo Yep. > But it seems to be trivial to spoof it > > login as: paul > paul@192.168.111.40's password: > Last login: Fri Apr 13 13:34:26 2007 from 192.168.111.101 > [paul@rhes5-wa-1 ~]$ sudo bash > [root@rhes5-wa-1 ~]# cat /proc/self/loginuid > 556[root@rhes5-wa-1 ~]# echo 600 > /proc/self/loginuid But you will generate an audit message right here saying that you changed the loginuid. While its not protection, it doesn't quietly let that occur. > I was 556 at login, after sudo i am still 556 but then as root I can now > change it to 600 and the audit log for my actions has auid=600 in it > doesnt that undermine the whole point of the login id? There should still be a trail showing that you changed it. > Surely once it has been set it should not be possible to change it again. I think there are 1 or 2 daemons that count on that. > I see a debate in the thread "proc_loginuid_write() checks wrong > capability" about who should be able to do this but it misses the point. It > should only be writtable if its -1 at the moment. Otherwise it must be > unconditionally rejected. Except in the case where daemons are changing loginuid temporarily to record the fact that its doing something on behalf of a user. For example, I think postfix may do this when delivering email and reading the dot files. > The argument for it being like it is is that root is all powerfull so they > should be allowed to do anything. But the beauty of auid is that it lets me > see who the root sudoer really is. But there is a problem here...if you strip the CAP_AUDIT_CONTROL bit and and admin does a restart of sshd, sshd will not be able to set it either. If you only allow setting auid if its non-zero, how do you restart a service like sshd since it will inherit the loginuid? The only option is to reboot the machine, so forget about rpm upgrades. I really think the answer is to protect this with SE Linux policy...but there is a problem. I think that by setting loginuid by the proc filesystem we won't be able to have a SE Linux type that is distinct for loginuid without changing things in the kernel. Then you also have to worry about descriptors being held open across forks. Not to mention, you have to do a open(), write(), close() for each setting of loginuid. So I think the best solution is to make a syscall for setting the loginuid that SE Linux can protect with a distinct type. Then we can say things like local_login_t is allowed setloginuid_t. > (BTW - my 10$ says its should be a new capability since it doesn't match > either CONTROL or WRITE, I've been saying that for a while too. I think it is distinct in the same way that chown is different than unlink. Changing the loginuid is only 1 attribute in the event. CAP_AUDIT_CONTROL on the otherhand allows you to delete rules. So you change whether or not the event even was recorded as opposed to who to blame. But one thing that does help is deleting the rules is done via netlink interface, which SE Linux does control. So, policy can be written that disallows anything except auditctl_t from writing to netlink, and policy can be written saying who can run auditctl. -Steve