Command vs. Protocol: Throughout this document, we will use bold text to refer to a command or application. This is used for instances such as ssh, since it is a protocol as well as command.
The sections that follow will cover the methods of securing your DragonFly system that were mentioned in the last section of this chapter.
First off, do not bother securing staff accounts if you have not secured the root account. Most systems have a password assigned to the root account. The first thing you do is assume that the password is always compromised. This does not mean that you should remove the password. The password is almost always necessary for console access to the machine. What it does mean is that you should not make it possible to use the password outside of the console or possibly even with the su(1) command. For example, make sure that your pty's are specified as being insecure in the /etc/ttys file so that direct root logins via telnet or rlogin are disallowed. If using other login services such as sshd, make sure that direct root logins are disabled there as well. You can do this by editing your /etc/ssh/sshd_config file, and making sure that PermitRootLogin is set to NO. Consider every access method -- services such as FTP often fall through the cracks. Direct root logins should only be allowed via the system console.
Of course, as a sysadmin you have to be able to get to root, so we open up a few holes. But we make sure these holes require additional password verification to operate. One way to make root accessible is to add appropriate staff accounts to the wheel group (in /etc/group). The staff members placed in the wheel group are allowed to su to root. You should never give staff members native wheel access by putting them in the wheel group in their password entry. Staff accounts should be placed in a staff group, and then added to the wheel group via the /etc/group file. Only those staff members who actually need to have root access should be placed in the wheel group. It is also possible, when using an authentication method such as Kerberos, to use Kerberos' .k5login file in the root account to allow a ksu(1) to root without having to place anyone at all in the wheel group. This may be the better solution since the wheel mechanism still allows an intruder to break root if the intruder has gotten hold of your password file and can break into a staff account. While having the wheel mechanism is better than having nothing at all, it is not necessarily the safest option.
An indirect way to secure staff accounts, and ultimately root access is to use an alternative login access method and do what is known as ``starring'' out the encrypted password for the staff accounts. Using the vipw(8) command, one can replace each instance of an encrypted password with a single ``*'' character. This command will update the /etc/master.passwd file and user/password database to disable password-authenticated logins.
A staff account entry such as:
foobar:R9DT/Fa1/LV9U:1000:1000::0:0:Foo Bar:/home/foobar:/usr/local/bin/tcsh
Should be changed to this:
foobar:*:1000:1000::0:0:Foo Bar:/home/foobar:/usr/local/bin/tcsh
This change will prevent normal logins from occurring, since the encrypted password will never match ``*''. With this done, staff members must use another mechanism to authenticate themselves such as kerberos(1) or ssh(1) using a public/private key pair. When using something like Kerberos, one generally must secure the machines which run the Kerberos servers and your desktop workstation. When using a public/private key pair with ssh, one must generally secure the machine used to login from (typically one's workstation). An additional layer of protection can be added to the key pair by password protecting the key pair when creating it with ssh-keygen(1). Being able to ``star'' out the passwords for staff accounts also guarantees that staff members can only login through secure access methods that you have set up. This forces all staff members to use secure, encrypted connections for all of their sessions, which closes an important hole used by many intruders: sniffing the network from an unrelated, less secure machine.
The more indirect security mechanisms also assume that you are logging in from a more restrictive server to a less restrictive server. For example, if your main box is running all sorts of servers, your workstation should not be running any. In order for your workstation to be reasonably secure you should run as few servers as possible, up to and including no servers at all, and you should run a password-protected screen blanker. Of course, given physical access to a workstation an attacker can break any sort of security you put on it. This is definitely a problem that you should consider, but you should also consider the fact that the vast majority of break-ins occur remotely, over a network, from people who do not have physical access to your workstation or servers.
Using something like Kerberos also gives you the ability to disable or change the password for a staff account in one place, and have it immediately affect all the machines on which the staff member may have an account. If a staff member's account gets compromised, the ability to instantly change his password on all machines should not be underrated. With discrete passwords, changing a password on N machines can be a mess. You can also impose re-passwording restrictions with Kerberos: not only can a Kerberos ticket be made to timeout after a while, but the Kerberos system can require that the user choose a new password after a certain period of time (say, once a month).
The prudent sysadmin only runs the servers he needs to, no more, no less. Be aware that third party servers are often the most bug-prone. For example, running an old version of imapd or popper is like giving a universal root ticket out to the entire world. Never run a server that you have not checked out carefully. Many servers do not need to be run as root. For example, the ntalk, comsat, and finger daemons can be run in special user sandboxes. A sandbox is not perfect, unless you go through a large amount of trouble, but the onion approach to security still stands: If someone is able to break in through a server running in a sandbox, they still have to break out of the sandbox. The more layers the attacker must break through, the lower the likelihood of his success. Root holes have historically been found in virtually every server ever run as root, including basic system servers. If you are running a machine through which people only login via sshd and never login via telnetd or rshd or rlogind, then turn off those services!
DragonFly now defaults to running ntalkd, comsat, and finger in a sandbox. Another program which may be a candidate for running in a sandbox is named(8). /etc/defaults/rc.conf includes the arguments necessary to run named in a sandbox in a commented-out form. Depending on whether you are installing a new system or upgrading an existing system, the special user accounts used by these sandboxes may not be installed. The prudent sysadmin would research and implement sandboxes for servers whenever possible.
There are a number of other servers that typically do not run in sandboxes: sendmail, popper, imapd, ftpd, and others. There are alternatives to some of these, but installing them may require more work than you are willing to perform (the convenience factor strikes again). You may have to run these servers as root and rely on other mechanisms to detect break-ins that might occur through them.
The other big potential root holes in a system are the suid-root and sgid binaries installed on the system. Most of these binaries, such as rlogin, reside in /bin, /sbin, /usr/bin, or /usr/sbin. While nothing is 100% safe, the system-default suid and sgid binaries can be considered reasonably safe. Still, root holes are occasionally found in these binaries. A root hole was found in Xlib in 1998 that made xterm (which is typically suid) vulnerable. It is better to be safe than sorry and the prudent sysadmin will restrict suid binaries, that only staff should run, to a special group that only staff can access, and get rid of (chmod 000) any suid binaries that nobody uses. A server with no display generally does not need an xterm binary. Sgid binaries can be almost as dangerous. If an intruder can break an sgid-kmem binary, the intruder might be able to read /dev/kmem and thus read the encrypted password file, potentially compromising any passworded account. Alternatively an intruder who breaks group kmem can monitor keystrokes sent through pty's, including pty's used by users who login through secure methods. An intruder that breaks the tty group can write to almost any user's tty. If a user is running a terminal program or emulator with a keyboard-simulation feature, the intruder can potentially generate a data stream that causes the user's terminal to echo a command, which is then run as that user.
User accounts are usually the most difficult to secure. While you can impose Draconian access restrictions on your staff and ``star'' out their passwords, you may not be able to do so with any general user accounts you might have. If you do have sufficient control, then you may win out and be able to secure the user accounts properly. If not, you simply have to be more vigilant in your monitoring of those accounts. Use of ssh and Kerberos for user accounts is more problematic, due to the extra administration and technical support required, but still a very good solution compared to a crypted password file.
The only sure fire way is to * out as many passwords as you can and use ssh or Kerberos for access to those accounts. Even though the encrypted password file (/etc/spwd.db) can only be read by root, it may be possible for an intruder to obtain read access to that file even if the attacker cannot obtain root-write access.
Your security scripts should always check for and report changes to the password file (see the Checking file integrity section below).
If an attacker breaks root he can do just about anything, but there are certain conveniences. For example, most modern kernels have a packet sniffing device driver built in. Under DragonFly it is called the bpf device. An intruder will commonly attempt to run a packet sniffer on a compromised machine. You do not need to give the intruder the capability and most systems do not have the need for the bpf device compiled in.
But even if you turn off the bpf
device, you still have
/dev/mem and
/dev/kmem
to worry about. For that matter, the intruder can still write to
raw disk devices. Also, there is another kernel feature called
the module loader, kldload(8). An enterprising intruder can
use a KLD module to install his own bpf
device, or other sniffing
device, on a running kernel. To avoid these problems you have to
run the kernel at a higher secure level, at least securelevel 1.
The securelevel can be set with a sysctl on
the kern.securelevel variable. Once you have
set the securelevel to 1, write access to raw devices will be
denied and special chflags flags,
such as schg,
will be enforced. You must also ensure that the
schg flag is set on critical startup binaries,
directories, and script files -- everything that gets run up
to the point where the securelevel is set. This might be overdoing
it, and upgrading the system is much more difficult when you
operate at a higher secure level. You may compromise and run the
system at a higher secure level but not set the
schg flag for every system file and directory
under the sun. Another possibility is to simply mount
/ and /usr read-only.
It should be noted that being too Draconian in what you attempt to
protect may prevent the all-important detection of an
intrusion.
When it comes right down to it, you can only protect your core system configuration and control files so much before the convenience factor rears its ugly head. For example, using chflags to set the schg bit on most of the files in / and /usr is probably counterproductive, because while it may protect the files, it also closes a detection window. The last layer of your security onion is perhaps the most important -- detection. The rest of your security is pretty much useless (or, worse, presents you with a false sense of safety) if you cannot detect potential incursions. Half the job of the onion is to slow down the attacker, rather than stop him, in order to give the detection side of the equation a chance to catch him in the act.
The best way to detect an incursion is to look for modified, missing, or unexpected files. The best way to look for modified files is from another (often centralized) limited-access system. Writing your security scripts on the extra-secure limited-access system makes them mostly invisible to potential attackers, and this is important. In order to take maximum advantage you generally have to give the limited-access box significant access to the other machines in the business, usually either by doing a read-only NFS export of the other machines to the limited-access box, or by setting up ssh key-pairs to allow the limited-access box to ssh to the other machines. Except for its network traffic, NFS is the least visible method -- allowing you to monitor the filesystems on each client box virtually undetected. If your limited-access server is connected to the client boxes through a switch, the NFS method is often the better choice. If your limited-access server is connected to the client boxes through a hub, or through several layers of routing, the NFS method may be too insecure (network-wise) and using ssh may be the better choice even with the audit-trail tracks that ssh lays.
Once you give a limited-access box, at least read access to the client systems it is supposed to monitor, you must write scripts to do the actual monitoring. Given an NFS mount, you can write scripts out of simple system utilities such as find(1) and md5(1). It is best to physically md5 the client-box files at least once a day, and to test control files such as those found in /etc and /usr/local/etc even more often. When mismatches are found, relative to the base md5 information the limited-access machine knows is valid, it should scream at a sysadmin to go check it out. A good security script will also check for inappropriate suid binaries and for new or deleted files on system partitions such as / and /usr.
When using ssh rather than NFS, writing the security script is much more difficult. You essentially have to scp the scripts to the client box in order to run them, making them visible, and for safety you also need to scp the binaries (such as find) that those scripts use. The ssh client on the client box may already be compromised. All in all, using ssh may be necessary when running over insecure links, but it is also a lot harder to deal with.
A good security script will also check for changes to user and staff members access configuration files: .rhosts, .shosts, .ssh/authorized_keys and so forth... files that might fall outside the purview of the MD5 check.
If you have a huge amount of user disk space, it may take too long to run through every file on those partitions. In this case, setting mount flags to disallow suid binaries and devices on those partitions is a good idea. The nodev and nosuid options (see mount(8)) are what you want to look into. You should probably scan them anyway, at least once a week, since the object of this layer is to detect a break-in whether or not the break-in is effective.
Process accounting (see accton(8)) is a relatively low-overhead feature of the operating system which might help as a post-break-in evaluation mechanism. It is especially useful in tracking down how an intruder has actually broken into a system, assuming the file is still intact after the break-in occurs.
Finally, security scripts should process the log files, and the logs themselves should be generated in as secure a manner as possible -- remote syslog can be very useful. An intruder tries to cover his tracks, and log files are critical to the sysadmin trying to track down the time and method of the initial break-in. One way to keep a permanent record of the log files is to run the system console to a serial port and collect the information on a continuing basis through a secure machine monitoring the consoles.
A little paranoia never hurts. As a rule, a sysadmin can add any number of security features, as long as they do not affect convenience, and can add security features that do affect convenience with some added thought. Even more importantly, a security administrator should mix it up a bit -- if you use recommendations such as those given by this document verbatim, you give away your methodologies to the prospective attacker who also has access to this document.
This section covers Denial of Service attacks. A DoS attack is typically a packet attack. While there is not much you can do about modern spoofed packet attacks that saturate your network, you can generally limit the damage by ensuring that the attacks cannot take down your servers.
Limiting server forks.
Limiting springboard attacks (ICMP response attacks, ping broadcast, etc.).
Kernel Route Cache.
A common DoS attack is against a forking server that attempts
to cause the server to eat processes, file descriptors, and memory,
until the machine dies. inetd
(see inetd(8)) has several
options to limit this sort of attack. It should be noted that
while it is possible to prevent a machine from going down, it is
not generally possible to prevent a service from being disrupted
by the attack. Read the inetd manual
page carefully and pay
specific attention to the -c, -C,
and -R options. Note that spoofed-IP attacks
will circumvent the -C option to
inetd, so
typically a combination of options must be used. Some standalone
servers have self-fork-limitation parameters.
Sendmail has its
-OMaxDaemonChildren option, which tends to work
much better than trying to use sendmail's load limiting options
due to the load lag. You should specify a
MaxDaemonChildren parameter, when you start
sendmail, high enough to handle your
expected load, but not so high that the computer cannot handle that
number of sendmails without falling on
its face. It is also prudent to run sendmail in queued mode
(-ODeliveryMode=queued) and to run the daemon
(sendmail -bd) separate from the queue-runs
(sendmail -q15m). If you still want real-time
delivery you can run the queue at a much lower interval, such as
-q1m, but be sure to specify a reasonable
MaxDaemonChildren option for
that sendmail to prevent cascade failures.
Syslogd can be attacked directly
and it is strongly recommended that you use the -s
option whenever possible, and the -a option
otherwise.
You should also be fairly careful with connect-back services such as tcpwrapper's reverse-identd, which can be attacked directly. You generally do not want to use the reverse-ident feature of tcpwrappers for this reason.
It is a very good idea to protect internal services from
external access by firewalling them off at your border routers.
The idea here is to prevent saturation attacks from outside your
LAN, not so much to protect internal services from network-based
root compromise.
Always configure an exclusive firewall, i.e.,
``firewall everything except ports A, B,
C, D, and M-Z''. This way you can firewall off all of your
low ports except for certain specific services such as
named (if you are primary for a zone),
ntalkd,
sendmail, and other Internet-accessible
services. If you try to configure the firewall the other way
-- as an inclusive or permissive firewall, there is a good
chance that you will forget to ``close'' a couple of
services, or that you will add a new internal service and forget
to update the firewall. You can still open up the high-numbered
port range on the firewall, to allow permissive-like operation,
without compromising your low ports. Also take note that DragonFly
allows you to control the range of port numbers used for dynamic
binding, via the various net.inet.ip.portrange
sysctl's (sysctl -a | fgrep
portrange), which can also ease the complexity of your
firewall's configuration. For example, you might use a normal
first/last range of 4000 to 5000, and a hiport range of 49152 to
65535, then block off everything under 4000 in your firewall
(except for certain specific Internet-accessible ports, of
course).
Another common DoS attack is called a springboard attack
-- to attack a server in a manner that causes the server to
generate responses which overloads the server, the local
network, or some other machine. The most common attack of this
nature is the ICMP ping broadcast attack.
The attacker spoofs ping packets sent to your LAN's broadcast
address with the source IP address set to the actual machine they
wish to attack. If your border routers are not configured to
stomp on ping's to broadcast addresses, your LAN winds up
generating sufficient responses to the spoofed source address to
saturate the victim, especially when the attacker uses the same
trick on several dozen broadcast addresses over several dozen
different networks at once. Broadcast attacks of over a hundred
and twenty megabits have been measured. A second common
springboard attack is against the ICMP error reporting system.
By constructing packets that generate ICMP error responses, an
attacker can saturate a server's incoming network and cause the
server to saturate its outgoing network with ICMP responses. This
type of attack can also crash the server by running it out of
mbuf's, especially if the server cannot drain the ICMP responses
it generates fast enough. The DragonFly kernel has a new kernel
compile option called ICMP_BANDLIM
which limits the effectiveness
of these sorts of attacks. The last major class of springboard
attacks is related to certain internal
inetd services such as the
udp echo service. An attacker simply spoofs a UDP packet with the
source address being server A's echo port, and the destination
address being server B's echo port, where server A and B are both
on your LAN. The two servers then bounce this one packet back and
forth between each other. The attacker can overload both servers
and their LANs simply by injecting a few packets in this manner.
Similar problems exist with the internal
chargen port. A
competent sysadmin will turn off all of these inetd-internal test
services.
Spoofed packet attacks may also be used to overload the kernel
route cache. Refer to the net.inet.ip.rtexpire,
rtminexpire, and rtmaxcache
sysctl parameters. A spoofed packet attack
that uses a random source IP will cause the kernel to generate a
temporary cached route in the route table, viewable with
netstat -rna | fgrep W3. These routes
typically timeout in 1600 seconds or so. If the kernel detects
that the cached route table has gotten too big it will dynamically
reduce the rtexpire but will never decrease it
to less than rtminexpire. There are two
problems:
The kernel does not react quickly enough when a lightly loaded server is suddenly attacked.
The rtminexpire is not low enough for
the kernel to survive a sustained attack.
If your servers are connected to the Internet via a T3 or
better, it may be prudent to manually override both
rtexpire and rtminexpire
via sysctl(8). Never set either parameter to zero (unless
you want to crash the machine). Setting both
parameters to two seconds should be sufficient to protect the route
table from attack.
There are a few issues with both Kerberos and
ssh that need to be addressed if
you intend to use them. Kerberos V is an excellent
authentication protocol, but there are bugs in the kerberized
telnet and
rlogin applications that make them
unsuitable for dealing with binary streams. Also, by default
Kerberos does not encrypt a session unless you use the
-x option. ssh
encrypts everything by default.
ssh works quite well in every respect except that it forwards encryption keys by default. What this means is that if you have a secure workstation holding keys that give you access to the rest of the system, and you ssh to an insecure machine, your keys are usable. The actual keys themselves are not exposed, but ssh installs a forwarding port for the duration of your login, and if an attacker has broken root on the insecure machine he can utilize that port to use your keys to gain access to any other machine that your keys unlock.
We recommend that you use ssh in combination with Kerberos whenever possible for staff logins. ssh can be compiled with Kerberos support. This reduces your reliance on potentially exposable ssh keys while at the same time protecting passwords via Kerberos. ssh keys should only be used for automated tasks from secure machines (something that Kerberos is unsuited to do). We also recommend that you either turn off key-forwarding in the ssh configuration, or that you make use of the from=IP/DOMAIN option that ssh allows in its authorized_keys file to make the key only usable to entities logging in from specific machines.
Contact the Documentation mailing list for comments, suggestions and questions about this document.