![]() | ![]() |
|
Security Quick-Start HOWTO for Linuxv. 1.2, 2002-07-21
1. Introduction1.1. Why me?Who should be reading this document and why should the average Linux user care about security? Those new to Linux, or unfamiliar with the inherent security issues of connecting a Linux system to large networks like Internet should be reading. "Security" is a broad subject with many facets, and is covered in much more depth in other documents, books, and on various sites on the Web. This document is intended to be an introduction to the most basic concepts as they relate to Linux, and as a starting point only.
The above is real, live data from a one week period for my home LAN. Much of the above would seem to be specifically targeted at Linux systems. Many of the targeted "destination" ports are used by well known Linux and Unix services, and all may be installed, and possibly even running, on your system. The focus here will be on threats that are shared by all Linux users, whether a dual boot home user, or large commercial site. And we will take a few, relatively quick and easy steps that will make a typical home Desktop system or small office system running Linux reasonably safe from the majority of outside threats. For those responsible for Linux systems in a larger or more complex environment, you'd be well advised to read this, and then follow up with additional reading suitable to your particular situation. Actually, this is probably good advice for everybody. We will assume the reader knows little about Linux, networking, TCP/IP, and the finer points of running a server Operating System like Linux. We will also assume, for the sake of this document, that all local users are "trusted" users, and won't address physical or local network security issues in any detail. Again, if this is not the case, further reading is strongly recommended. The principles that will guide us in our quest are:
The steps we will be taking to get there are:
If you don't have time to read everything, concentrate on Steps 1, 2, and 3. This is where the meat of the subject matter is. The Appendix has a lot of supporting information, which may be helpful, but may not be necessary for all readers. 1.2. CopyrightSecurity-Quickstart HOWTO for Linux Copyright © 2001 Hal Burgiss. This document is free; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 2 of the License, or (at your option) any later version. This document is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. You can get a copy of the GNU GPL at at http://www.gnu.org/copyleft/gpl.html. 1.3. CreditsMany thanks to those who helped with the production of this document.
1.4. DisclaimerThe author accepts no liability for the contents of this document. Use the concepts, examples and other content at your own risk. As this is a new document, there may be errors and inaccuracies. Hopefully these are few and far between. Corrections and suggestions are welcomed. This document is intended to give the new user a starting point for securing their system while it is connected to the Internet. Please understand that there is no intention whatsoever of claiming that the contents of this document will necessarily result in an ultimately secure and worry-free computing environment. Security is a complex topic. This document just addresses some of the most basic issues that inexperienced users should be aware of. The reader is encouraged to read other security related documentation and articles. And to stay abreast of security issues as they evolve. Security is not an objective, but an ongoing process. 1.5. New Versions and ChangelogThe current official version can always be found at http://www.tldp.org/HOWTO/Security-Quickstart-HOWTO/. Pre-release versions can be found at http://feenix.burgiss.net/ldp/quickstart/. Other formats, including PDF, PS, single page HTML, may be found at the Linux Documentation HOWTO index page: http://tldp.org/docs.html#howto. Changelog: Version 1.2: Clarifications on example firewall scripts, and small additions to 'Have I been Hacked'. Note on Zonealarm type applications. More on the use of "chattr" by script kiddies, and how to check for this. Other small additions and clarifications. Version 1.1: Various corrections, amplifications and numerous mostly small additions. Too many to list. Oh yea, learn to spell Red Hat correctly ;-) Version 1.0: This is the initial release of this document. Comments welcomed. 1.6. FeedbackAny and all comments on this document are most welcomed. Please make sure you have the most current version before submitting corrections or suggestions! These can be sent to <hal@foobox.net>. 2. ForewordBefore getting into specifics, let's try to briefly answer some questions about why we need to be concerned about security in the first place. It is easy to see why an e-commerce site, an on-line bank, or a government agency with sensitive documents would be concerned about security. But what about the average user? Why should even a Linux home Desktop user worry about security? Anyone connected to the Internet is a target, plain and simple. It makes little difference whether you have a part-time dialup connection, or a full-time connection, though full-time connections make for bigger targets. Larger sites make for bigger targets too, but this does not let small users off the hook since the "small user" may be less skilled and thus an easier victim. There are those out there that are scanning just for easy victims all the time. If you start logging unwanted connection attempts, you will see this soon enough. There is little doubt that many of these attempts are maliciously motivated and the attacker, in some cases, is looking for Linux boxes to crack. Does someone on the other side of the globe really want to borrow my printer? What do they want? Often, they just may want your computer, your IP address, and your bandwidth. Then they use you to either attack others, or possibly commit crimes or mischief and are hiding their true identity behind you. This is an all too common scenario. Commercial and high-profile sites are targeted more directly and have bigger worries, but we all face this type of common threat. With a few reasonable precautions, Linux can be very secure, and with all the available tools, makes for a fantastically fun and powerful Internet connection or server. Most successful break-ins are the result of ignorance or carelessness. The bottom line is:
These are all real possibilities, unless we take the appropriate precautions.
2.1. The Optimum ConfigurationIdeally, we would want one computer as a dedicated firewall and router. This would be a bare bones installation, with no servers running, and only the required services and components installed. The rest of our systems would connect via this dedicated router/firewall system. If we wanted publicly accessible servers (web, mail, etc), these would be in a "DMZ" (De-militarized Zone). The router/firewall allows connections from outside to whatever services are running in the DMZ by "forwarding" these requests, but it is segregated from the rest of the internal network (aka LAN) otherwise. This leaves the rest of the internal network in fairly secure isolation, and relative safety. The "danger zone" is confined to the DMZ. But not everyone has the hardware to dedicate to this kind of installation. This would require a minimum of two computers. Or three, if you would be running any publicly available servers (not a good idea initially). Or maybe you are just new to Linux, and don't know your way around well enough yet. So if we can't do the ideal installation, we will do the next best thing. 2.2. Before We StartBefore we get to the actual configuration sections, a couple of notes. First, one of the interesting aspects of Linux, is the different distributions like Caldera, Red Hat, SuSE, and Debian. While these are all "Linux", and may share certain features, there is surely some differences as to what utilities they may install as defaults. Most Linux distributions will write their own system configuration tools as well. And with Linux, there is always more than one way to skin a cat. But for the purposes of our discussion, we will have to use as generic set of tools as we can. Unfortunately, GUI tools don't lend themselves to this type of documentation. We will be using text based, command line tools for the most part. If you are familiar with your distribution's utilities, feel free to substitute those in appropriate places. And if not, you should learn them or suitable alternatives. The next several sections have been written such that you can perform the recommended procedures as you read along. This is the "Quick Start" in the document title! To get ready, what you will need for the configuration sections below:
We'll be using a hypothetical system here for examples with the hostname "bigcat". Bigcat is a Linux desktop with a fresh install of the latest/greatest Linux distro running. Bigcat has a full-time, direct Internet connection. Even if your installation is not so "fresh", don't be deterred. Better late than never. 3. Step 1: Which services do we really need?In this section we will see which services are running on our freshly installed system, decide which we really need, and do away with the rest. If you are not familiar with how servers and TCP connections work, you may want to read the section on servers and ports in the Appendix first. If not familiar with the netstat utility, you may want to read a quick overview of it beforehand. There is also a section in the Appendix on ports, and corresponding services. You may want to look that over too. Our goal is to turn off as many services as possible. If we can turn them all off, or at least off to outside connections, so much the better. Some rules of thumb we will use to guide us:
3.1. System AuditSo what is really running on our system anyway? Let's not take anything for granted about what "should" be running, or what we "think" is running. Unfortunately, there is no such things as a standard Linux installation. The wide variety of servers available, coupled with each particular distribution's installation options, make providing a ready made list impossible. The best that can be done is show you how to list all running services, and point you in the right general direction. Now open an xterm, and su to root. You'll need to widen the window wide so the lines do not wrap. Use this command: netstat -tap |grep LISTEN. This will give us a list of all currently running servers as indicated by the keyword LISTEN, along with the "PID" and "Program Name" that started each particular service.
Note the first three columns are cropped above for readability. If your list is as long as the example, you have some work ahead of you! It is highly unlikely that you really need anywhere near this number of servers running. Please be aware that the example above is just one of many, many possible system configurations. Yours probably does look very different. You don't understand what any of this is telling you? Hopefully then, you've read the netstat tutorial in the Appendix, and understand how it works. Understanding exactly what each server is in the above example, and what it does, is beyond the scope of this document. You will have to check your system's documentation (e.g. Installation Guide, man pages, etc) if that service is important to you. For example, does "exec", "login", and "shell" sound important? Yes, but these are not what they may sound like. They are actually rexec, rlogin, and rsh, the "r" (for remote) commands. These are antiquated, unnecessary, and in fact, are very dangerous if exposed to the Internet. Let's make a few quick assumptions about what is necessary and unnecessary, and therefore what goes and what stays on bigcat. Since we are running a desktop on bigcat, X11 of course needs to stay. If bigcat were a dedicated server of some kind, then X11 would be unnecessary. If there is a printer physically attached, the printer (lp) daemon should stay. Otherwise, it goes. Print servers may sound harmless, but are potential targets too since they can hold ports open. If we plan on logging in to bigcat from other hosts, sshd (Secure SHell Daemon) would be necessary. If we have Microsoft hosts on our LAN, we probably want Samba, so smbd should stay. Otherwise, it is completely unnecessary. Everything else in this example is optional and not required for a normally functioning system, and should probably go. See anything that you don't recognize? Not sure about? It goes! To sum up: since bigcat is a desktop with a printer attached, we will need "x11", "printer". bigcat is on a LAN with MS hosts, and shares files and printing with them, so "netbios-ssn" (smbd) is desired. We will also need "ssh" so we can login from other machines. Everything else is unnecessary for this particular case. Nervous about this? If you want, you can make notes of any changes you make or save the list of servers you got from netstat, with this command: netstat -tap |grep LISTEN > ~/services.lst. That will save it your home directory with the name of "services.lst" for future reference. This is to not say that the ones we have decided to keep are inherently safe. Just that we probably need these. So we will have to deal with these via firewalling or other means (addressed below). It is worth noting that the telnet and ftp daemons in the above example are servers, aka "listeners". These accept incoming connections to you. You do not need, or want, these just to use ftp or telnet clients. For instance, you can download files from an FTP site with just an ftp client. Running an ftp server on your end is not required at all, and has serious security implications. There may be individual situations where it is desirable to make exceptions to the conclusions reached above. See below. 3.2. The Danger Zone (or r00t m3 pl34s3)The following is a list of services that should not be run over the Internet. Either disable these (see below), uninstall, or if you really do need these services running locally, make sure they are the current, patched versions and that they are effectively firewalled. And if you don't have a firewall in place now, turn them off until it is up and verified to be working properly. These are potentially insecure by their very nature, and as such are prime cracker targets.
This is not necessarily a definitive list. Just some common services that are sometimes started on default Linux installations. And conversely, this does not imply that other services are inherently safe. 3.3. Stopping ServicesThe next step is to find where each server on our kill list is being started. If it is not obvious from the netstat output, use ps, find, grep or locate to find more information from the "Program name" or "PID" info in the last column. There is examples of this in the Process Owner section in the netstat Tutorial of the Appendix. If the service name or port number do not look familiar to you, you might get a real brief explanation in your /etc/services file. Skeptical that we are going to break your system, and the pieces won't go back together again? If so, take this approach: turn off everything listed above in "The Danger Zone", and run your system for a while. OK? Try stopping one of the ones we found to be "unnecessary" above. Then, run the system for a while. Keep repeating this process, until you get to the bare minimum. If this works, then make the changes permanent (see below). The ultimate objective is not just to stop the service now, but to make sure it is stopped permanently! So whatever steps you take here, be sure to check after your next reboot. There are various places and ways to start system services. Let's look at the most common ways this is done, and is probably how your system works. System services are typically either started by "init" scripts, or by inetd (or its replacement xinetd) on most distributions. (The location of the init scripts may vary from distribution to distribution.) 3.3.1. Stopping Init ServicesInit services are typically started automatically during the boot process, or during a runlevel change. There is a naming scheme that uses symlinks to determine which services are to be started, or stopped, at any given runlevel. The scripts themselves should be in /etc/init.d/ (or possibly /etc/rc.d/init.d/ ). This init style is used by Red Hat, SuSE, Mandrake, Debian, Conectiva, and most Linuxes. Slackware is one notable exception (though recent versions have an option for this)! Typically on Slackware system services are all configured in one file: /etc/rc.d/rc.inet2. You can get a listing of these scripts:
Or use whichever tools your distribution provides for this. To stop a running service now, as root (on SysVinit style systems, which is pretty much everybody):
Where "$SERVICE_NAME" is the name of the init script, which is often, but not always, the same as the service name itself. This should do the trick on most distributions. Older Red Hat versions may use the path /etc/rc.d/init.d/ instead. This only stops this particular service now. It will restart again on the next reboot, or runlevel change, unless additional steps are taken. So this is really a two step process for init type services. Your distribution will have utilities available for controlling which services are started at various runlevels. Debian based systems have update-rc.d for this, and Red Hat based systems have chkconfig. If you are familiar with these tools, do it now, and then check again after the next reboot. If you are not familiar with these tools, see the man pages and learn it now! This is something that you need to know. For Debian (where $SERVICE_NAME is the init script name):
And Red Hat:
Another option here is to uninstall a package if you know you do not need it. This is a pretty sure-fire, permanent fix. This also alleviates the potential problem of keeping all installed packages updated and current (Step 2). And, package management systems like RPM or DEB make it very easy to re-install a package should you change your mind. 3.3.2. InetdInetd is called a "super-daemon" because it is used to spawn sub-daemons. inetd itself will generally be started via init scripts, and will "listen" on the various ports as determined by which services are enable in its configuration file, /etc/inetd.conf. Any service listed here will be under the control of inetd. Likewise, any of the listening servers in netstat output that list "inetd" in the last column under "Program Name", will have been started by inetd. You will have to adjust the inetd configuration to stop these services. xinetd is an enhanced inetd replacement, and is configured differently (see next section below). Below is a partial snippet from a typical inetd.conf. Any service with a "#" at the beginning of the line is "commented out", and thus ignored by inetd, and consequently disabled.
The above example has two services enabled: time and pop3. To disable these, all we need is to open the file with a text editor, comment out the two services with a "#", save the file, and then restart inetd (as root):
Check your logs for errors, and run netstat again to verify all went well. A quicker way of getting the same information, using grep:
Again, do you see anything there that you don't know what it is? Then in all likelihood you are not using it, and it should be disabled. Unlike the init services configuration, this is a lasting change so only the one step is required. Let's expose one myth that gets tossed around: you shouldn't disable a service by commenting out, or removing, entries from /etc/services. This may have the desired effect in some cases, but is not the right way to do it, and may interfere with the normal operation of other system utilities. 3.3.3. Xinetdxinetd is an inetd replacement with enhancements. It essentially serves the same purpose as inetd, but the configuration is different. The configuration can be in the file /etc/xinetd.conf, or individual files in the directory /etc/xinetd.d/. Turning off xinetd services is done by either deleting the corresponding configuration section, or file. Or by using your text editor and simply setting disable = yes for the appropriate service. Then, xinetd will need to be restarted. See man xinetd and man xinetd.conf for syntax and configuration options. A sample xinetd configuration:
You can get a quick list of enabled services:
At this point, the above output should raise some red flags. In the overwhelming majority of systems, all the above can be disabled without any adverse impact. Not sure? Try it without that service. After disabling unnecessary services, then restart xinetd:
3.3.4. When All Else FailsOK, if you can't find the "right" way to stop a service, or maybe a service is being started and you can't find how or where, you can "kill" the process. To do this, you will need to know the PID (Process I.D.). This can be found with ps, top, fuser or other system utilities. For top and ps, this will be the number in the first column. See the Port and Process Owner section in the Appendix for examples. Example (as root):
Then run top or ps again to verify that the process is gone. If not, then:
Note the second "KILL" in there. This must be done either by the user who owns the process, or root. Now go find where and how this process got started ;-) The /proc filesystem can also be used to find out more information about each process. Armed with the PID, we can find the path to a mysterious process:
3.4. ExceptionsAbove we used the criteria of turning off all unnecessary services. Sometimes that is not so obvious. And sometimes what may be required for one person's configuration is not the same for another's. Let's look at a few common services that fall in this category. Again, our rule of thumb is if we don't need it, we won't run it. It's that simple. If we do need any of these, they are prime candidates for some kind of restrictive policies via firewall rules or other mechanisms (see below).
3.5. Summary and Conclusions for Step 1In this section we learned how to identify which services are running on our system, and were given some tips on how to determine which services may be necessary. Then we learned how to find where the services were being started, and how to stop them. If this has not made sense, now is a good time to re-read the above. Hopefully you've already taken the above steps. Be sure to test your results with netstat again, just to verify the desired end has been achieved, and only the services that are really required are running. It would also be wise to do this after the next reboot, anytime you upgrade a package (to make sure a new configuration does not sneak in), and after every system upgrade or new install. 4. Step 2: UpdatingOK, this section should be comparatively short, simple and straightforward compared to the above, but no less important. The very first thing after a new install you should check your distribution's updates and security notices and apply all patches . Only a year old you say? That's a long time actually, and not current enough to be safe. Only a few months or few weeks? Check anyway. A day or two? Better safe than sorry. It is quite possible that security updates have been released during the pre-release phase of the development and release cycle. If you can't take this step, disable any publicly accessible services until you can. Linux distributions are not static entities. They are updated with new, patched packages as the need arises. The updates are just as important as the original installation. Even more so, since they are fixes. Sometimes these updates are bug fixes, but quite often they are security fixes because some hole has been discovered. Such "holes" are immediately known to the cracker community, and they are quick to exploit them on a large scale. Once the hole is known, it is quite simple to get in through it, and there will be many out there looking for it. And Linux developers are also equally quick to provide fixes. Sometimes the same day as the hole has become known! Keeping all installed packages current with your release is one of the most important steps you can take in maintaining a secure system. It can not be emphasized enough that all installed packages should be kept updated -- not just the ones you use. If this is burdensome, consider uninstalling any unused packages. Actually this is a good idea anyway. But where to get this information in a timely fashion? There are a number of web sites that offer the latest security news. There are also a number of mailing lists dedicated to this topic. In fact, your vendor most likely has such a list where vulnerabilities and the corresponding fix is announced. This is an excellent way to stay abreast of issues effecting your release, and is highly recommended. http://linuxsecurity.com is a good site for Linux only issues. They also have weekly newsletters available: http://www.linuxsecurity.com/general/newsletter.html. Also, many distributions have utilities that will automatically update your installed packages via ftp. This can be run as a cron job on a regular basis and is a painless way to go if you have ready Internet access. This is not a one time process -- it is ongoing. It is important to stay current. So watch those security notices. And subscribe to your vendor's security mailing list today! If you have cable modem, DSL, or other full time connection, there is no excuse not to do this religiously. All distributions make this easy enough! One last note: any time a new package is installed, there is also a chance that a new or revised configuration has been installed as well. Which means that if this package is a server of some kind, it may be enabled as a result of the update. This is bad manners, but it can happen, so be sure to run netstat or comparable to verify your system is where you want it after any updates or system changes. In fact, do it periodically even if there are no such changes. 4.1. Summary and Conclusions for Step 2It is very simple: make sure your Linux installation is current. Check with your vendor for what updated packages may be available. There is nothing wrong with running an older release, just so the packages in it are updated according to what your vendor has made available since the initial release. At least as long as your vendor is still supporting the release and updates are still being provided. 5. Step 3: Firewalls and Setting Access PoliciesSo what is a "firewall"? It's a vague term that can mean anything that acts as a protective barrier between us and the outside world. This can be a dedicated system, or a specific application that provides this functionality. Or it can be a combination of components, including various combinations of hardware and software. Firewalls are built from "rules" that are used to define what is allowed to enter and exit a given system or network. Let's look at some of the possible components that are readily available for Linux, and how we might implement a reasonably safe firewalling strategy. In Step 1 above, we have turned off all services we don't need. In our example, there were a few we still needed to have running. In this section, we will take the next step here and decide which we need to leave open to the world. And which we might be able to restrict in some way. If we can block them all, so much the better, but this is not always practical. 5.1. StrategyWhat we want to do now is restrict connections and traffic so that we only allow the minimum necessary for whatever our particular situation is. In some cases we may want to block all incoming "new" connection attempts. Example: we want to run X, but don't want anyone from outside to access it, so we'll block it completely from outside connections. In other situations, we may want to limit, or restrict, incoming connections to trusted sources only. The more restrictive, the better. Example: we want to ssh into our system from outside, but we only ever do this from our workplace. So we'll limit sshd connections to our workplace address range. There are various ways to do this, and we'll look at the most common ones. We also will not want to limit our firewall to any one application. There is nothing wrong with a "layered" defense-in-depth approach. Our front line protection will be a packet filter -- either ipchains or iptables (see below). Then we can use additional tools and mechanisms to reinforce our firewall. We will include some brief examples. Our rule of thumb will be to deny everything as the default policy, then open up just what we need. We'll try to keep this as simple as possible since it can be an involved and complex topic, and just stick to some of the most basic concepts. See the Links section for further reading on this topic. 5.2. Packet Filters -- Ipchains and Iptables"Packet filters" (like ipchains) have the ability to look at individual packets, and make decisions based on what they find. These can be used for many purposes. One common purpose is to implement a firewall. Common packet filters on Linux are ipchains which is standard with 2.2 kernels, and iptables which is available with the more recent 2.4 kernels. iptables has more advanced packet filtering capabilities and is recommended for anyone running a 2.4 kernel. But either can be effective for our purposes. ipfwadm is a similar utility for 2.0 kernels (not discussed here). If constructing your own ipchains or iptables firewall rules seems a bit daunting, there are various sites that can automate the process. See the Links section. Also the included examples may be used as a starting point. And your distribution may be including a utility of some kind for generating a firewall script. This may be adequate, but it is still recommended to know the proper syntax and how the various mechanisms work as such tools rarely do more than a few very simple rules.
5.2.1. ipchainsipchains can be used with either 2.2 or 2.4 kernels. When ipchains is in place, it checks every packet that moves through the system. The packets move across different "chains", depending where they originate and where they are going. Think of "chains" as rule sets. In advanced configurations, we could define our own custom chains. The three default built-in chains are input, which is incoming traffic, output, which is outgoing traffic, and forward, which is traffic being forwarded from one interface to another (typically used for "masquerading"). Chains can be manipulated in various ways to control the flow of traffic in and out of our system. Rules can be added at our discretion to achieve the desired result. At the end of every "chain" is a "target". The target is specified with the -j option to the command. The target is what decides the fate of the packet and essentially terminates that particular chain. The most common targets are mostly self-explanatory: ACCEPT, DENY, REJECT, and MASQ. MASQ is for "ipmasquerading". DENY and REJECT essentially do the same thing, though in different ways. Is one better than the other? That is the subject of much debate, and depends on other factors that are beyond the scope of this document. For our purposes, either should suffice. ipchains has a very flexible configuration. Port (or port ranges), interfaces, destination address, source address can be specified, as well as various other options. The man page explains these details well enough that we won't get into specifics here. Traffic entering our system from the Internet, enters via the input chain. This is the one that we need as tight as we can make it. Below is a brief example script for a hypothetical system. We'll let the comments explain what this script does. Anything starting with a "#" is a comment. ipchains rules are generally incorporated into shell scripts, using shell variables to help implement the firewalling logic.
To use the above script would require that it is executable (i.e. chmod +x ipchains.sh), and run by root to build the chains, and hence the firewall. To summarize what this example did was to start by setting some shell variables in the top section, to be used later in the script. Then we set the default rules (ipchains calls these "policies") of denying all inbound and forwarded traffic, and of allowing all our own outbound traffic. We had to open some holes in the high, unprivileged ports so that we could have return traffic from connections that bigcat initiates to outside addresses. If we connect to someone's web server, we want that HTML data to be able to get back to us, for instance. The same applies to other network traffic. We then allowed a few specific types of the ICMP protocol (most are still blocked). We are also logging any inbound traffic that violates any of our rules so we know who is doing what. Notice that we are only using IP address here, not hostnames of any kind. This is so that our firewall works, even in situation where there may be DNS failures. Also, to prevent any kind of DNS spoofing. See the ipchains man page for a full explanation of syntax. The important ones we used here are:
By and large, the order in which command line options are specified is not significant. The chain name (e.g. input) must come first though. Remember in Step 1 when we ran netstat, we had both X and print servers running among other things. We don't want these exposed to the Internet, even in a limited way. These are still happily running on bigcat, but are now safe and sound behind our ipchains based firewall. You probably have other services that fall in this category as well. The above example is a simplistic all or none approach. We allow all our own outbound traffic (not necessarily a good idea), and block all inbound connection attempts from outside. It is only protecting one interface, and really just the inbound side of that interface. It would more than likely require a bit of fine tuning to make it work for you. For a more advanced set of rules, see the Appendix. And you might want to read http://tldp.org/HOWTO/IPCHAINS-HOWTO.html. Whenever you have made changes to your firewall, you should verify its integrity. One step to make sure your rules seem to be doing what you intended, is to see how ipchains has interpreted your script. You can do this by opening your xterm very wide, and issuing the following command:
The output is grouped according to chain. You should also find a way to scan yourself (see the Verifying section below). And then keep an eye on your logs to make sure you are blocking what is intended. 5.2.2. iptablesiptables is the next generation packet filter for Linux, and requires a 2.4 kernel. It can do everything ipchains can, but has a number of noteworthy enhancements. The syntax is similar to ipchains in many respects. See the man page for details. The most noteworthy enhancement is "connection tracking", also known as "stateful inspection". This gives iptables more knowledge of the state of each packet. Not only does it know if the packet is a TCP or UDP packet, or whether it has the SYN or ACK flags set, but also if it is part of an existing connection, or related somehow to an existing connection. The implications for firewalling should be obvious. The bottom line is that it is easier to get a tight firewall with iptables, than with ipchains. So this is the recommended way to go. Here is the same script as above, revised for iptables:
The same script logic is used here, and thus this does pretty much the same exact thing as the ipchains script in the previous section. There are some subtle differences as to syntax. Note the case difference in the chain names for one (e.g. INPUT vs input). Logging is handled differently too. It has its own "target" now (-j LOG), and is much more flexible. There are some very fundamental differences as well, that might not be so obvious. Remember this section from the ipchains script:
We jumped through hoops here with ipchains so that we could restrict unwanted, incoming connections as much as possible. A bit of a kludge, actually. That section is missing from the iptables version. It is not needed as connection tracking handles this quite nicely, and then some. This is due to the "statefulness" of iptables. It knows more about each packet than ipchains. For instance, it knows whether the packet is part of a "new" connection, or an "established" connection, or a "related" connection. This is the so-called "stateful inspection" of connection tracking. There are many, many features of iptables that are not touched on here. For more reading on the Netfilter project and iptables, see http://netfilter.samba.org. And for a more advanced set of rules, see the Appendix. 5.3. Tcpwrappers (libwrap)Tcpwrappers provides much the same desired results as ipchains and iptables above, though works quite differently. Tcpwrappers actually intercepts the connection attempt, then examines its configurations files, and decides whether to accept or reject the request. Tcpwrappers controls access at the application level, rather than the socket level like iptables and ipchains. This can be quite effective, and is a standard component on most Linux systems. Tcpwrappers consists of the configuration files /etc/hosts.allow and /etc/hosts.deny. The functionality is provided by the libwrap library. Tcpwrappers first looks to see if access is permitted in /etc/hosts.allow, and if so, access is granted. If not in /etc/hosts.allow, the file /etc/hosts.deny is then checked to see if access is not allowed. If so, access is denied. Else, access is granted. For this reason, /etc/hosts.deny should contain only one uncommented line, and that is: ALL: ALL. Access should then be permitted through entries in /etc/hosts.allow, where specific services are listed, along with the specific host addresses allowed to access these services. While hostnames can be used here, use of hostnames opens the limited possibility for name spoofing. Tcpwrappers is commonly used to protect services that are started via inetd (or xinetd). But also any program that has been compiled with libwrap support, can take advantage of it. Just don't assume that all programs have built in libwrap support -- they do not. In fact, most probably don't. So we will only use it in our examples here to protect services start via inetd. And then rely on our packet filtering firewall, or other mechanism, to protect non-(x)inetd services. Below is a small snippet from a typical inetd.conf file:
The second to last column is the tcpwrappers daemon -- /usr/sbin/tcpd. Immediately after is the daemon it is protecting. In this case, POP and IMAP mail servers. Your distro probably has already done this part for you. For the few applications that have built-in support for tcpwrappers via the libwrap library, specifying the daemon as above is not necessary. We will use the same principles here: default policy is to deny everything, then open holes to allow the minimal amount of traffic necessary. So now with your text editor, su to root and open /etc/hosts.deny. If it does not exist, then create it. It is just a plain text file. We want the following line:
If it is there already, fine. If not, add it in and then save and close file. Easy enough. "ALL" is one of the keywords that tcpwrappers understands. The format is $SERVICE_NAME : $WHO, so we are denying all connections to all services here. At least all services that are using tcpwrappers. Remember, this will primarily be inetd services. See man 5 hosts_access for details on the syntax of these files. Note the "5" there! Now let's open up just the services we need, as restrictively as we can, with a brief example:
The first line allows all "localhost" connections. You will need this. The second allows connections to the sshd and ipop3d services from IP addresses that start with 192.168.1., in this case the private address range for our hypothetical home LAN. Note the trailing ".". It's important. The third line allows connections to only our sshd daemon from any host associated with .myworkplace.com. Note the leading "." in this example. And then also, the single host hostess.mymomshouse.com. In summary, localhost and all our LAN connections have access to any and all tcpwrappered services on bigcat. But only our workplace addresses, and our mother can use sshd on bigcat from outside connections. Everybody else is denied by the default policy in /etc/hosts.deny. The types of wild cards above (.myworkplace.com and 192.168.1.) are not supported by ipchains and iptables, or most other Linux applications for that matter. Also, tcpwrappers can use hostnames in place of IP addresses which is quite handy in some situations. This does not work with ipchains and iptables. You can test your tcpwrappers configuration with the included tcpdchk utility (see the man page). Note that at this time this does not work with xinetd, and may not even be included in this case. There is nothing wrong with using both tcpwrappers and a packet filtering firewall like ipchains. In fact, it is recommended to use a "layered" approach. This helps guard against accidental misconfigurations. In this case, each connection will be tested by the packet filter rules first, then tcpwrappers. Remember to make backup copies before editing system configuration files, restart the daemon afterward, and then check the logs for error messages. 5.3.1. xinetdAs mentioned, xinetd is an enhanced inetd . It has much of the same functionality, with some notable enhancements. One is that tcpwrappers support can be compiled in, eliminating the need for explicit references to tcpd. Which means /etc/hosts.allow and /etc/hosts.deny are automatically in effect. Don't assume this is the case though. A little testing, then viewing the logs should be able to tell you whether tcpwrappers support is automatic or not. Some of xinetd's other enhancements: specify IP address to listen on, which is a very effective method of access control; limit the rate of incoming connections and the total number of simultaneous connections; limit services to specific times of day. See the xinetd and xinetd.conf man pages for more details. The syntax is quite different though. An example from /etc/xinetd.d/tftp:
Notice the bind statement. We are only listening on, or "binding" to, the private, LAN interface here. No outside connections can be made since the outside port is not even opened. We are also only accepting connections from 192.168.1.0, our LAN. For xinetd's purposes, this denotes any IP address beginning with "192.168.1". Note that the syntax is different from inetd. The server statement in this case is the tftp daemon, in.tftpd. Again, this assumes that libwrap/tcpwrappers support is compiled into xinetd. The user running the daemon will be "nobody". Yes, there is a user account called "nobody", and it is wise to run such daemons as non-root users whenever possible. Lastly, the disable statement is xinetd's way of turning services on or off. In this case, it is "on". This is on here only as an example. Do NOT run tftp as a public service as it is unsafe. 5.4. PortSentryPortsentry works quite differently than the other tools discussed so far. Portsentry does what its name implies -- it guards ports. Portsentry is configured with the /etc/portsentry/portsentry.conf file. Unlike the other applications discussed above, it does this by actually becoming the listening server on those ports. Kind of like baiting a trap. Running netstat -taup as root while portsentry is running, will show portsentry as the LISTENER on whatever ports portsentry is configured for. If portsentry senses a connection attempt, it blocks it completely. And then goes a step further and blocks the route to that host to stop all further traffic. Alternately, ipchains or iptables can be used to block the host completely. So it makes an excellent tool to stop port scanning of a range of ports. But portsentry has limited flexibility as to whether it allows a given connection. It is pretty much all or nothing. You can define specific IP addresses that it will ignore in /etc/portsentry/portsentry.ignore. But you cannot allow selective access to individual ports. This is because only one server can bind to a particular port at the same time, and in this case that is portsentry itself. So it has limited usefulness as a stand-alone firewall. As part of an overall firewall strategy, yes, it can be quite useful. For most of us, it should not be our first line of defense, and we should only use it in conjunction with other tools. Suggestion on when portsentry might be useful:
All in all, the packet filters make for a better firewall. 5.5. ProxiesThe dictionary defines "proxy" as "the authority or power to act on behalf of another". This pretty well describes software proxies as well. It is an intermediary in the connection path. As an example, if we were using a web proxy like "squid" (http://www.squid-cache.org/), every time we browse to a web site, we would actually be connecting to our locally running squid server. Squid in turn, would relay our request to the ultimate, real destination. And then squid would relay the web pages back to us. It is a go-between. Like "firewalls", a "proxy" can refer to either a specific application, or a dedicated server which runs a proxy application. Proxies can perform various duties, not all of which have much to do with security. But the fact that they are an intermediary, makes them a good place to enforce access control policies, limit direct connections through a firewall, and control how the network behind the proxy looks to the Internet. So this makes them strong candidates to be part of an overall firewall strategy. And, in fact, are sometimes used instead of packet filtering firewalls. Proxy based firewalls probably make more sense where many users are behind the same firewall. And it probably is not high on the list of components necessary for home based systems. Configuring and administering proxies can be complex, and is beyond the scope of this document. The Firewall and Proxy Server HOWTO, http://tldp.org/HOWTO/Firewall-HOWTO.html, has examples of setting up proxy firewalls. Squid usage is discussed at http://squid-docs.sourceforge.net/latest/html/book1.htm 5.6. Individual ApplicationsSome servers may have their own access control features. You should check this for each server application you run. We'll only look at a few of the common ones in this section. Man pages, and other application specific documentation, is your friend here. This should be done whether you have confidence in your firewall or not. Again, layers of protection is always best.
As always, anytime you make system changes, backup the configuration file first, restart the appropriate daemon afterward, and then check the appropriate logs for error messages. 5.7. VerifyingThe final step after getting your firewall in place, is to verify that it is doing what you intended. You would be wise to do this anytime you make even minor changes to your system configuration. So how to do this? There are several things you can do. For our packet filters like ipchains and iptables, we can list all our rules, chains, and associated activity with iptables -nvL | less (substitute ipchains if appropriate). Open your xterm as wide as possible to avoid wrapping long lines. This should give you an idea if your chains are doing what you think they should. You may want to perform some of the on-line tasks you normally do first: open a few web pages, send and retrieve mail, etc. This will, of course, not give you any information on tcpwrappers or portsentry. tcpdchk can be used to verify tcpwrappers configuration (except with xinetd). And then, scan yourself. nmap is the scanning tool of choice and may be available via your distribution , or from http://www.insecure.org/nmap/nmap_download.html. nmap is very flexible, and essentially is a "port prober". In other words, it looks for open ports, among other things. See the nmap man page for details. If you do run nmap against yourself (e.g. nmap localhost), this should tell you what ports are open -- and visible locally only! Which hopefully by now, is quite different from what can be seen from the outside. So, scan yourself, and then find a trusted friend, or site (see the Links section), to scan you from the outside. Make sure you are not violating your ISPs Terms of Service by port scanning. It may not be allowed, even if the intentions are honorable. Scanning from outside is the best way to know how the rest of the world sees you. This should tell you how well that firewall is working. See the nmap section in the Appendix for some examples on nmap usage. One caveat on this: some ISPs may filter some ports, and you will not know for sure how well your firewall is working. Conversely, they make it look like certain ports are open by using web, or other, proxies. The scanner may see the web proxy at port 80 and mis-report it as an open port on your system. Another option is to find a website that offers full range testing. http://www.hackerwhacker.com is one such site. Make sure that any such site is not just scanning a relatively few well known ports. Repeat this procedure with every firewall change, every system upgrade or new install, and when any key components of your system changes. You may also want to enable logging all the denied traffic. At least temporarily. Once the firewall is verified to be doing what you think it should, and if the logs are hopelessly overwhelming, you may want to disable logging. If relying on portsentry at all, please read the documentation. Depending on your configuration it will either drop the route to the scanner, or implement a ipchains/iptables rule doing the same thing. Also, since it "listens" on the specified ports, all those ports will show as "open". A false alarm in this case. 5.8. LoggingLinux does a lot of logging. Usually to more than one file. It is not always obvious what to make of all these entries -- good, bad or indifferent? Firewall logs tend to generate a fair amount of each. Of course, you are wanting to stop only the "bad", but you will undoubtedly catch some harmless traffic as well. The 'net has a lot of background noise. In many cases, knowing the intentions of an incoming packet are almost impossible. Attempted intrusion? Misbehaved protocol? Mis-typed IP address? Conclusions can be drawn based on factors such as destination port, source port, protocol, and many other variables. But there is no substitute for experience in interpreting firewall logs. It is a black art in many cases. So do we really need to log? And how much should we be trying to log? Logging is good in that it tells us that the firewall is functional. Even if we don't understand much of it, we know it is doing "something". And if we have to, we can dig into those logs and find whatever data might be called for. On the other hand, logging can be bad if it is so excessive, it is difficult to find pertinent data, or worse, fills up a partition. Or if we over re-act and take every last entry as an all out assault. Some perspective is a great benefit, but something that new users lack almost by definition. Again, once your firewall is verified, and you are perplexed or overwhelmed, home desktop users may want to disable as much logging as possible. Anyone with greater responsibilities should log, and then find ways to extract the pertinent data from the logs by filtering out extraneous information. Not sure where to look for log data? This could conceivably be many places depending on how your distribution configured the various daemons and syslogd. Most logging is done in /var/log/*. Check that directory with ls -l /var/log/ and see if you can tell the most active logs by size and timestamp. Also, look at /etc/syslog.conf to see where the default logs are. /var/log/messages is a good place to look for starters. Portsentry and tcpwrappers do a certain amount of logging that is not adjustable. xinetd has logging enhancements that can be turned on. Both ipchains and iptables, on the other hand, are very flexible as to what is logged. For ipchains the -l option can be added to any rule. iptables uses the -j LOG target, and requires its own, separate rule instead. iptables goes a few steps further and allows customized log entries, and rate limiting. See the man page. Presumably, we are more interested in logging blocked traffic, so we'd confine logging to only our DENY and REJECT rules. So whether you log, and how much you log, and what you do with the logs, is an individual decision, and probably will require some trial and error so that it is manageable. A few auditing and analytical tools can be quite helpful: Some tools that will monitor your logs for you and notify you when necessary. These likely will require some configuration, and trial and error, to make the most out of them:
5.9. Where to StartLet's take a quick look at where to run our firewall scripts from. Portsentry can be run as an init process, like other system services. It is not so important when this is done. Tcpwrappers will be automatically be invoked by inetd or xinetd, so not to worry there either. But the packet filtering scripts will have to be started somewhere. And many scripts will have logic that uses the local IP address. This will mean that the script must be started after the interface has come up and been assigned an IP address. Ideally, this should be immediately after the interface is up. So this depends on how you connect to the Internet. Also, for protocols like PPP or DHCP that may be dynamic, and get different IP's on each re-connect, it is best to have the scripts run by the appropriate daemon. For PPP, you probably have an /etc/ppp/ip-up file. This will be executed every time there is a connect or re-connect. You should put the full path to your firewall script here. Check the local documentation for the correct location. Debian use files in /etc/ppp/ip-up.d/, so either put the script itself there, or a symlink to it. Red Hat uses /etc/ppp/ip-up.local for any user defined, local PPP configuration. For DHCP, it depends on which client. dhcpcd will execute /etc/dhcpcd/dhcpcd-<interface>.exe (e.g. dhcpcd-eth0.exe) whenever a lease is obtained or renewed. So this is where to put a reference to your firewall script. For pump, the main configuration file is /etc/pump.conf. Pump will run whatever script is defined by the "script" statement any time there is a new or renewed lease:
If you have a static IP address (i.e. it never changes), the placement is not so important and should be before the interface comes up! 5.10. Summary and Conclusions for Step 3In this section we looked at various components that might be used to construct a "firewall". And learned that a firewall is as much a strategy and combination of components, as it is any one particular application or component. We looked at a few of the most commonly available applications that can be found on most, if not all, Linux systems. This is not a definitive list. This is a lot of information to digest at all at one time and expect anyone to understand it all. Hopefully this can used as a starting point, and used for future reference as well. The packet filter firewall examples can be used as starting points as well. Just use your text editor, cut and paste into a file with an appropriate name, and then run chmod +x against it to make it executable. Some minor editing of the variables may be necessary. Also look at the Links section for sites and utilities that can be used to generate a custom script. This may be a little less daunting. Now we are done with Steps 1, 2 and 3. Hopefully by now you have already instituted some basic measures to protect your system(s) from the various and sundry threats that lurk on networks. If you haven't implemented any of the above steps yet, now is a good time to take a break, go back to the top, and have at it. The most important steps are the ones above. A few quick conclusions... "What is best iptables, ipchains, tcpwrappers, or portsentry?" The quick answer is that iptables can do more than any of the others. So if you are using a 2.4 kernel, use iptables. Then, ipchains if using a 2.2 kernel. The long answer is "it just depends on what you are doing and what the objective is". Sorry. The other tools all have some merit in any given situation, and all can be effective in the right situation. "Do I really need all these packages?" No, but please combine more than one approach, and please follow all the above recommendations. iptables by itself is good, but in conjunction with some of the other approaches, we are even stronger. Do not rely on any single mechanism to provide a security blanket. "Layers" of protection is always best. As is sound administrative practices. The best iptables script in the world is but one piece of the puzzle, and should not be used to hide other system weaknesses. "If I have a small home LAN, do I need to have a firewall on each computer?" No, not necessary as long as the LAN gateway has a properly configured firewall. Unwanted traffic should be stopped at that point. And as long as this is working as intended, there should be no unwanted traffic on the LAN. But, by the same token, doing this certainly does no harm. And on larger LANs that might be mixed platform, or with untrusted users, it would be advisable. 6. Intrusion DetectionThis section will deal with how to get early warning, how to be alerted after the fact, and how to clean up from intrusion attempts. 6.1. Intrusion Detection Systems (IDS)Intrusion Detection Systems (IDS for short) are designed to catch what might have gotten past the firewall. They can either be designed to catch an active break-in attempt in progress, or to detect a successful break-in after the fact. In the latter case, it is too late to prevent any damage, but at least we have early awareness of a problem. There are two basic types of IDS: those protecting networks, and those protecting individual hosts. For host based IDS, this is done with utilities that monitor the filesystem for changes. System files that have changed in some way, but should not change -- unless we did it -- are a dead give away that something is amiss. Anyone who gets in, and gets root, will presumably make changes to the system somewhere. This is usually the very first thing done. Either so he can get back in through a backdoor, or to launch an attack against someone else. In which case, he has to change or add files to the system. This is where tools like tripwire (http://www.tripwire.org) play a role. Such tools monitor various aspects of the filesystem, and compare them against a stored database. And can be configured to send an alert if any changes are detected. Such tools should only be installed on a known "clean" system. For home desktops and home LANs, this is probably not an absolutely necessary component of an overall security strategy. But it does give peace of mind, and certainly does have its place. So as to priorities, make sure the Steps 1, 2 and 3 above are implemented and verified to be sound, before delving into this. RPM users can get somewhat the same results with rpm -Va, which will verify all packages, but without all the same functionality. For instance, it will not notice new files added to most directories. Nor will it detect files that have had the extended attributes changed (e.g. chattr +i, man chattr and man lsattr). For this to be helpful, it needs to be done after a clean install, and then each time any packages are upgraded or added. Example:
Then we have a stored system snapshot that we can refer back to. Debian users have a similar tool with debsums.
Another idea is to run chkrootkit (http://www.chkrootkit.org/) as a weekly cron job. This will detect common "rootkits". 6.2. Have I Been Hacked?Maybe you are reading this because you've noticed something "odd" about your system, and are suspicious that someone was gotten in? This can be a clue. The first thing an intruder typically does is install a "rootkit". There are many prepackaged rootkits available on the Internet. The rootkit is essentially a script, or set of scripts, that makes quick work of modifying the system so the intruder is in control, and he is well hidden. He does this by installing modified binaries of common system utilities and tampering with log files. Or by using special kernel modules that achieve similar results. So common commands like ls may be modified so as to not show where he has his files stored. Clever! A well designed rootkit can be quite effective. Nothing on the system can really be trusted to provide accurate feedback. Nothing! But sometimes the modifications are not as smooth as intended and give hints that something is not right. Some things that might be warning signs:
Sometimes the intruder is not so smart and forgets about root's .bash_history, or cleaning up log entries, or even leaves strange, leftover files in /tmp. So these should always be checked too. Just don't necessarily expect them to be accurate. Often such left behind files, or log entries, will have obvious script kiddie sounding names, e.g. "r00t.sh". Packet sniffers, like tcpdump (http://www.tcpdump.org), might be useful in finding any uninvited traffic. Interpreting sniffer output is probably beyond the grasp of the average new user. snort (http://www.snort.org), and ethereal (http://www.ethereal.com), are also good. Ethereal has a GUI. As mentioned, a compromised system will undoubtedly have altered system binaries, and the output of system utilities is not to be trusted. Nothing on the system can be relied upon to be telling you the whole truth. Re-installing individual packages may or may not help since it could be system libraries or kernel modules that are doing the dirty work. The point here is that there is no way to know with absolute certainty exactly what components have been altered. RPM users can use rpm -Va |less to attempt to verify the integrity all packages. But again there is no assurance that rpm itself has not been tampered with, or the system components that RPM relies on. If you have pstree on your system, try this instead of the standard ps. Sometimes the script kiddies forget about this one. No guarantees though that this is accurate either. You can also try querying the /proc filesystem, which contains everything the kernel knows about processes that are running:
This will provide a list of all processes and PID numbers (assuming a malicious kernel module is not hiding this). Another approach is to visit http://www.chkrootkit.org, download their rootkit checker, and see what it says. Some interesting discussions on issues surrounding forensics can be found at http://www.fish.com/security/. There is also a collection of tools available, aptly called "The Coroner's Toolkit" (TCT). Read below for steps on recovering from an intrusion. 6.3. Reclaiming a Compromised SystemSo now you've confirmed a break-in, and know that someone else has root access, and quite likely one or more hidden backdoors to your system. You've lost control. How to clean up and regain control? There is no sure fire way of doing this short of a complete re-install. There is no way to find with assurance all the modified files and backdoors that may have been left. Trying to patch up a compromised system risks a false sense of security and may actually aggravate an already bad situation. The steps to take, in this order:
At this time, any rootkit cleanup tools that may be available on-line are not recommended. They probably do work just fine most of the time. But again, how to be absolutely sure that all is well and all vestiges of the intrusion are gone? 7. General TipsThis section will quickly address some general concepts for maintaining a more secure and reliable system or network. Let's emphasize "maintaining" here since computer systems change daily, as does the environment around them. As mentioned before, there isn't any one thing that makes a system secure. There are too many variables. Security is an approach and an attitude more than it is a reliance on any particular product, application or specific policy.
8. Appendix8.1. Servers, Ports, and PacketsLet's take a quick, non-technical look at some networking concepts, and how they can potentially impact our own security. We don't need to know much about networking, but a general idea of how things work is certainly going to help us with firewalls and other related issues. As you may have noticed Linux is a very network oriented Operating System. Much is done by connecting to "servers" of one type or another -- X servers, font servers, print servers, etc. Servers provide "services", which provide various capabilities, both to the local system and potentially other remote systems. The same server generally provides both functionalities. Some servers work quietly behind the scenes, and others are more interactive by nature. We may only be aware of a print server when we need to print something, but it is there running, listening, and waiting for connection requests whether we ever use it or not (assuming of course we have it enabled). A typical Linux installation will have many, many types of servers available to it. Default installations often will turn some of these "on". And even if we are not connected to a real network all the time, we are still "networked" so to speak. Take our friendly local X server for instance. We may tend to think of this as just providing a GUI interface, which is only true to a point. It does this by "serving" to client applications, and thus is truly a server. But X Windows is also capable of serving remote clients over a network -- even large networks like the Internet. Though we probably don't really want to be doing this ;-) And yes, if you are not running a firewall or have not taken other precautions, and are connected to the Internet, it is quite possible that someone -- anyone -- could connect to your X server. X11 "listens" on TCP "port" 6000 by default. This principle applies to most other servers as well -- they can be easily connected to, unless something is done to restrict or prevent connections. In TCP/IP (Transmission Control Protocol/Internet Protocol) networks like we are talking about with Linux and the Internet, every connected computer has a unique "IP Address". Think of this like a phone number. You have a phone number, and in order to call someone else, you have to know that phone number, and then dial it. The phone numbers have to be unique for the system to work. IP address are generally expressed as "dotted quad" notation, e.g. 152.19.254.81. On this type of network, servers are said to "listen". This means that they have a "port" opened, and are awaiting incoming connections to that port. Connections may be local, as is typically the case with our X server, or remote -- meaning from another computer "somewhere". So servers "listen" on a specific "port" for incoming connections. Most servers have a default port, such as port 80 for web servers. Or 6000 for X11. See /etc/services for a list of common ports and their associated service. The "port" is actually just an address in the kernel's networking stack, and is a method that TCP, and other protocols, use to organize connections and the exchange of data between computers. There are total of 65,536 TCP and UDP ports available, though usually only a relatively few of these are used at any one time. These are classified as "privileged", those ports below 1024, and "unprivileged", 1024 and above. Most servers use the privileged ports. Only one server may listen on, or "bind" to, a port at a time. Though that server may well be able to open multiple connections via that one port. Computers talk to each other via these "port" connections. One computer will open a connection to a "port" on another computer, and thus be able to exchange data via the connection that has been established between their respective ports. Getting back to the phone analogy, and stretching it a bit, think of calling a large organization with a complex phone system. The organization has many "departments": sales, shipping, billing, receiving, customer service, R&D, etc. Each department has it's own "extension" number. So the shipping department might be extension 21, the sales might be department 80 and so on. The main phone number is the IP Address, and the department's extension is the port in this analogy. The "department's" number is always the same when we call. And generally they can handle many simultaneous incoming calls. The data itself is contained in "packets", which are small chunks of data, generally 1500 bytes or less each. Packets are used to control and organize the connection, as well as carry data. There are different types of packets. Some are specifically used for controlling the connection, and then some packets carry our data as their payload. If there is a lot of data, it will be broken up into multiple packets which is almost always how it works. The packets will be transmitted one at a time, and then "re-assembled" at the other end. One web page for instance, will take many packets to transmit -- maybe hundreds or even thousands. This all happens very quickly and transparently. We can see a typical connection between two computers in this one line excerpt from netstat output:
The interesting part is the IP addresses and ports in the fourth and fifth columns. The port is the number just to the right of the colon. The left side of the colon is the IP address of each computer. The fourth column is the local address, or our end of the connection. In the example, 169.254.179.139 is the IP address assigned by my ISP. It is connected to port 21 (FTP) on 18.29.1.67, which is rpmfind.net. This is just after an FTP download from rpmfind.net. Note that while I am connected to their FTP server on their port 21, the port on my end that is used by my FTP client is 1359. This is a randomly assigned "unprivileged" port, used for my end of the two-way "conversation". The data moves in both directions: me:port#1359 <-> them:port#21. The FTP protocol is actually a little more complicated than this, but we won't delve into the finer points here. The CLOSE_WAIT is the TCP state of the connection at this particular point in time. Eventually the connection will close completely on both ends, and netstat will not show anything for this. The "unprivileged" port that is used for my end of the connection, is temporary and is not associated with a locally running server. It will be closed by the kernel when the connection is terminated. This is quite different than the ports that are kept open by "listening" servers, which are permanent and remain "open" even after a remote connection is terminated. So to summarize using the above example, we have client (me) connecting to a server (rpmfind.net), and the connection is defined and controlled by the respective ports on either end. The data is transmitted and controlled by packets. The server is using a "privileged" port (i.e. a port below number 1024) which stays open listening for connections. The "unprivileged" port used on my end by my client application is temporary, is only opened for the duration of the connection, and only responds to the server's port at the other end of the connection. This type of port is not vulnerable to attacks or break-ins generally speaking. The server's port is vulnerable since it remains open. The administrator of the FTP server will need to take appropriate precautions that his server is secure. Other Internet connections, such as to web servers or mail servers, work similar to the above example, though the server ports are different. SMTP mail servers use port 25, and web servers typically use port 80. See the Ports section for other commonly used ports and services. One more point on ports: ports are only accessible if there is something listening on that port. No one can force a port open if there is no service or daemon listening there, ready to handle incoming connection requests. A closed port is a totally safe port. And a final point on the distinction between clients and servers. The example above did not have a telnet or ftp server in the LISTENER section in the netstat example above. In other words, no such servers were running locally. You do not need to run a telnet or ftp server daemon in order to connect to somebody else's telnet or ftp server. These are only for providing these services to others that would be making connections to you. Which you don't really want to be doing in most cases. This in no way effects the ability to use telnet and ftp client software. 8.2. Common PortsA quick run down of some commonly seen and used ports, with the commonly associated service name, and risk factor. All have some risk. It is just that some have historically had more exploits than others. That is how they are evaluated below, and not necessarily to be interpreted as whether any given service is safe or not.
More services and corresponding port numbers can be found in /etc/services. Also, the "official" list is http://www.iana.org/assignments/port-numbers. A great analysis of what probes to these and other ports might mean from Robert Graham: http://www.linuxsecurity.com/resource_files/firewalls/firewall-seen.html. A very good reference. Another point here, these are the standard port designations. There is no law that says any service has to run on a specific port. Usually they do, but certainly they don't always have to. Just a reminder that when you see these types of ports in your firewall logs, it is not anything to go off the deep end about. Not if you have followed Steps 1-3 above, and verified your firewall works. You are fairly safe. Much of this traffic may be "stray bullets" too -- Internet background noise, misconfigured clients or routers, noisy Windows stuff, etc. 8.3. Netstat Tutorial8.3.1. Overviewnetstat is a very useful utility for viewing the current state of your network status -- what servers are listening for incoming connections, what interfaces they listen on, who is connected to us, who we are connect to, and so on. Take a look at the man page for some of the many command line options. We'll just use a relative few options here. As an example, let's check all currently listening servers and active connections for both TCP and UDP on our hypothetical host, bigcat. bigcat is a home desktop installation, with a DSL Internet connection in this example. bigcat has two ethernet cards: one for the external connection to the ISP, and one for a small LAN with an address of 192.168.1.1.
This output probably looks very different from what you get on your own system. Notice the distinction between "Local Address" and "Foreign Address", and how each includes a corresponding port number (or service name if available) after the colon. "Local Address" is our end of the connection. The first group with LISTEN in the far right hand column are services that are running on this system. These are servers that are running in the background on bigcat, and "listen" for incoming connections. So they have a port opened, and this is where they "listen". These connections might come from the local system (i.e. bigcat itself), or remote systems. This is very important information to have! The others just below this are connections that have been established from this system to other systems. The respective connections are in varying states as indicated by the key words in the last column. Those with no key word in the last column at the end are servers responding to UDP connections. UDP is a different protocol from TCP altogether, but is used for some types of low priority network traffic. Now, the same thing with the "-n" flag to suppress converting to "names" so we can actually see the port numbers:
Let's look at the first few lines of this in detail. On line one,
"Local Address" is 0.0.0.0, meaning "all" interfaces that are available. The local port is 515, or the standard print server port, usually owned by the lpd daemon. You can find a listing of common service names and corresponding ports in the file /etc/services. The fact that it is listening on all interfaces is significant. In this case, that would be lo (localhost), eth0, and eth1. Printer connections could conceivably be made over any of these interfaces. Should a user on this system bring up a PPP connection, then the print daemon would be listening on that interface (ppp0) as well. The "Foreign Address" is also 0.0.0.0, meaning from "anywhere". It is also worth noting here, that even though this server is telling the kernel to listen on all interfaces, the netstat output does not reflect whether there may be a firewall in place that may be filtering incoming connections. We just can't tell that at this point. Obviously, for certain servers, this is very desirable. Nobody outside your own LAN has any reason whatsoever to connect to your print server port for instance. Line two is a little different:
Notice the "Local Address" this time is localhost's address of 127.0.0.1. This is very significant as only connections local to this machine will be accepted. So only bigcat can connect to bigcat's TCP port 8000. The security implications should be obvious. Not all servers have configuration options that allow this kind of restriction, but it is a very useful feature for those that do. Port 8000 in this example, is owned by the web proxy Junkbuster. With the next three entries, we are back to listening on all available interfaces:
Looking at /etc/services, we can tell that port 37 is a "time" service, which is a time server. 6000 is X11, and 80 is the standard port for HTTP servers like Apache. There is nothing really unusual here as these are all readily available services on Linux. The first two above are definitely not the kind of services you'd want just anyone to connect to. These should be firewalled so that all outside connections are refused. Again, we can't tell from this output whether any firewall is in place, much less how effectively implemented it may be. The web server on port 80 is not a huge security risk by itself. HTTP is a protocol that is often open to all comers. For instance, if we wanted to host our own home page, Apache can certainly do this for us. It is also possible to firewall this off, so that it is for use only to our LAN clients as part of an Intranet. Obviously too, if you do not have a good justification for running a web server, then it should be disabled completely. The next two lines are interesting:
Again notice the "Local Address" is not 0.0.0.0. This is good! The port this time is 53, or the DNS port used by nameserver daemons like named. But we see the nameserver daemon is only listening on the lo interface (localhost), and the interface that connects bigcat to the LAN. So the kernel only allows connections from localhost, and the LAN. There will be no port 53 available to outside connections at all. This is a good example of how individual applications can sometimes be securely configured. In this case, we are probably looking at a caching DNS server since a real nameserver that is responsible for handling DNS queries would have to have port 53 open to the world. This is a security risk and requires special handling. The last three LISTENER entries:
These are back to listening on all available interfaces. Port 22 is sshd, the Secure Shell server daemon. This is a good sign! Notice that the service for port 631 does not have a service name if we look at the output in the first example. This might be a clue that something unusual is going on here. (See the next section for the answer to this riddle.) And lastly, port 25, the standard port for the SMTP mail daemon. Most Linux installations probably will have an SMTP daemon running, so this is not necessarily unusual. But is it necessary? The next grouping is established connections. For our purposes the state of the connection as indicated by the last column is not so important. This is well explained in the man page.
There are nine total connections here. The first three is our external interface connecting to a remote host on their port 119, the standard NNTP (News) port. There are three connections here to the same news server. Apparently the application is multi-threaded, as it is trying to open multiple connections to the news server. The next two entries are connections to a remote web server as indicated by the port 80 after the colon in the fifth column. Probably a pretty common looking entry for most of us. But the one just after is reversed and has the port 80 in the fourth column, so this is someone that has connected to bigcat's web server via its external, Internet-side interface. The last three entries are all connections from localhost to localhost. So we are connecting to ourselves here. Remembering from above that port 8000 is bigcat's web proxy, this is a web browser that is connected to the locally running proxy. The proxy then will open an external connection of its own, which probably is what is going on with lines four and five. Since we gave netstat both the -t and -u options, we are getting both the TCP and UDP listening servers. The last few lines are the UDP ones:
The last three entries have ports that are familiar from the above discussion. These are servers that are listening for both TCP and UDP connections. Same servers in this case, just using two different protocols. The first one on local port 32768 is new, and does not have a service name available to it in /etc/services. So at first glance this should be suspicious and pique our curiosity. See the next section for the explanation. Can we draw any conclusions from this hypothetical situation? For the most part, these look to be pretty normal looking network services and connections for Linux. There does not seem to be an unduly high number of servers running here, but that by itself does not mean much since we don't know if all these servers are really required or not. We know that netstat can not tell us if any of these are effectively firewalled, so there is no way to say how secure all this might be. We also don't really know if all the listening services are really required by the owner here. That is something that varies widely from installation to installation. Does bigcat even have a printer attached for instance? Presumably it does, or this is a completely unnecessary risk. 8.3.2. Port and Process OwnersWe've learned a lot about what is going on with bigcat's networking from the above section. But suppose we see something we don't recognize and want to know what started that particular service? Or we want to stop a particular server and it is not obvious from the above output? The -p option should give us the process's PID and the program name that started the process in the last column. Let's look at the TCP servers again (with first three columns cropped for spacing). We'll have to run this as root to get all the available information:
Some of these we already know about. But we see now that the printer daemon on port 515 is being started via inetd with a PID of "988". inetd is a special situation. inetd is often called the "super server", since it's main role is to spawn sub-services. If we look at the first line, inetd is listening on port 515 for printer services. If a connection comes for this port, inetd intercepts it, and then will spawn the appropriate daemon, i.e. the print daemon in this case. The configuration of how inetd handles this is typically done in /etc/inetd.conf. This should tell us that if we want to stop an inetd controlled server on a permanent basis, then we will have to dig into the inetd (or perhaps xinetd) configuration. Also the time service above is started via inetd as well. This should also tell us that these two services can be further protected by tcpwrappers (discussed in Step 3 above). This is one benefit of using inetd to control certain system services. We weren't sure about the service on port 631 above since it did not have a standard service name, which means it is something maybe unusual or off the beaten path. Now we see it is owned by cupsd , which is one of several print daemons available under Linux. This happens to be the web interface for controlling the printer service. Something cupsd does that is indeed a little different than other print servers. The last entry above is the SMTP mail server on bigcat. Often, this is sendmail with many distributions. But not in this case. The command is "master", which may not ring any bells. Armed with the program name we could go searching the filesystem with tools like the locate or find commands. After we found it, we could then probably discern what package it belonged to. But with the PID available now, we can look at ps output, and see if that helps us any:
We took a shortcut here by combining ps with grep. It looks like that this file belongs to postfix, which is indeed a mail server package comparable to sendmail. Running ps with the --forest flag (-f for short) can be helpful in determining what processes are parent or child process or another process. An edited example:
A couple of things to note here. We have two by now familiar daemons here: named and postfix (smtpd). Both are spawning sub-processes. In the case of named, what we are seeing is threads, various sub-processes that it always spawns. Postfix is also spawning sub-processes, but not as "threads". Each sub-process has its own specific task. It is worth noting that child processes are dependent on the parent process. So killing the parent PID, will in turn kill all child processes. If all this has not shed any light, we might also try locate:
find is perhaps the most flexible file finding utility, but doesn't use a database the way locate does, so is much slower:
If lsof is installed, it is another command that is useful for finding who owns processes or ports:
This is again telling us that the cupsd print daemon is the owner of port 631. Just a different way of getting at it. Yet one more way to get at this is with fuser, which should be installed:
See the man pages for fuser and lsof command syntax. Another place to look for where a service is started, is in the init.d directory, where the actual init scripts live (for SysVinit systems). Something like ls -l /etc/init.d/, should give us a list of these. Often the script name itself gives a hint as to which service(s) it starts, though it may not necessarily exactly match the "Program Name" as provided by netstat. Or we can use grep to search inside files and match a search pattern. Need to find where rpc.statd is being started, and we don't see a script by this name?
We didn't really need all that information, but at least we see now exactly which script is starting it. Remember too that not all services are started this way. Some may be started via inetd, or xinetd. The /proc filesystem also keeps everything we want to know about processes that are running. We can query this to find out more information about each process. Do you need to know the full path of the command that started a process?
Finally, we had a loose end or two in the UDP listening services. Remember we had a strange looking port number 32768, that also had no service name associated with it:
Now by including the "PID/Program name" option with the -p flag, we see this also belongs to named, the nameserver daemon. Recent versions of BIND use an unprivileged port for some types of traffic. In this case, this is BIND 9.x. So no real alarms here either. The unprivileged port here is the one named uses to to talk to other nameservers for name and address lookups, and should not be firewalled. So we found no big surprises in this hypothetical situation. If all else fails, and you can't find a process owner for an open port, suspect that it may be an RPC (Remote Procedure Call) service of some kind. These use randomly assigned ports without any seeming logic or consistency, and are typically controlled by the portmap daemon. In some cases, these may not reveal the process owner to netstat or lsof. Try stopping portmap, and then see if the mystery service goes away. Or you can use rpcinfo -p localhost to see what RPC services may be running (portmap must be running for this to work).
8.4. Attacks and ThreatsIn this section, we will take a quick look at some of the common threats and techniques that are out there, and attempt to put them into some perspective. The corporate world, government agencies and high profile Internet sites have to be concerned with a much more diverse and challenging set of threats than the typical home desktop user. There are many reasons someone may want to break in to someone else's computer. It may be just for kicks, or any number of malicious reasons. They may just want a base from which to attack someone else. This is a very common motivation. The most common "attack" for most of us is from already compromised systems. The Internet is littered with computers that have been broken into, and are now doing their master's bidding blindly, in zombie-like fashion. They are programmed to scan massively large address ranges, probing each individual IP address as they go. Looking for one or more open ports, and then probing for known weaknesses if they get the chance. Very impersonal. Very methodical. And very effective. We are all in the path of such robotic scans. All because those responsible for these systems fail to do what you are doing now - taking steps to protect their system(s), and avoid being r00ted. These scans do not look at login banners that may be presented on connection. It will do little good to change your /etc/issue.net to pretend that you are running some obscure operating system. If they find something listening, they will try all of the exploits appropriate to that port, without regard to any indications your system may give. If it works, they are in -- if not, they will move on. 8.4.1. Port Scans and ProbesFirst, let's define "scan" and "probe" since these terms come up quite a bit. A "probe" implies testing if a given port is open or closed, and possibly what might be listening on that port. A "scan" implies either "probing" multiple ports on one or more systems. Or individual ports on multiple systems. So you might "scan" all ports on your own system for instance. Or a cracker might "scan" the 216.78.*.* address range to see who has port 111 open. Black hats can use scan and probe information to know what services are running on a given system, and then they might know what exploits to try. They may even be able to tell what Operating System is running, and even kernel version, and thus get even more information. "Worms", on the other hand, are automated and scan blindly, generally just looking for open ports, and then a susceptible victim. They are not trying to "learn" anything, the way a cracker might. The distinction between "scan" and "probe"is often blurred. Both can used in good ways, or in bad ways, depending on who is doing it, and why. You might ask a friend to scan you, for instance, to see how well your firewall is working. This is a legitimate use of scanning tools such as nmap. But what if someone you don't know does this? What is their intent? If it's your ISP, they may be trying to enforce their Terms of Service Agreement. Or maybe, it is someone just playing, and seeing who is "out there". But more than likely it is someone or something with not such good intentions. Full range port scans (meaning probing of many ports on the same machine) seem to be a not so common threat for home based networks. But certainly, scanning individual ports across numerous systems is a very, very common occurrence. 8.4.2. RootkitsA "rootkit" is the script kiddie's stock in trade. When a successful intrusion takes place, the first thing that is often done, is to download and install such "rootkits". The rootkit is a set of scripts designed to take control of the system, and then hide the intrusion. Rootkits are readily available on the web for various Operating Systems. A rootkit will typically replace critical system files such as ls, ps, netstat, login and others. Passwords may be added, hidden daemons started, logs tampered with, and surely one of more backdoors are opened. The hidden backdoors allow easy access any time the attacker wants back in. And often the vulnerability itself may even be fixed so that the new "owner" has the system all to himself. The entire process is scripted so it happens very quickly. The rightful owners of these compromised systems generally have no idea what is going on, and are victims themselves. A well designed rootkit can be very difficult to detect. 8.4.3. Worms and ZombiesA "worm" is a self replicating exploit. It infects a system, then attempts to spread itself typically via the same vulnerability. Various "worms" are weaving their way through the entire Internet address space constantly, spreading themselves as they go. But somewhere behind the zombie, there is a controller. Someone launched the worm, and they will be informed after a successful intrusion. It is then up to them how the system will be used. Many of these are Linux systems, looking for other Linux systems to "infect" via a number of exploits. But most Operating Systems share in this threat. Once a vulnerable system is found, the actual entry and take over is quick, and may be difficult to detect after the fact. The first thing an intruder (whether human or "worm") will do is attempt to cover their tracks. A "rootkit" is downloaded and installed. This trend has been exacerbated by the growing popularity of cable modems and DSL. The number of full time Internet connections is growing rapidly, and this makes fertile ground for such exploits since often these aren't as well secured as larger sites. While this may sound ominous, a few simple precautions can effectively deter this type of attack. With so many easy victims out there, why waste much effort breaking into your system? There is no incentive to really try very hard. Just scan, look, try, move on if unsuccessful. There is always more IPs to be scanned. If your firewall is effectively bouncing this kind of thing, it is no threat to you at all. Take comfort in that, and don't over re-act. It is worth noting, that these worms cannot "force" their way in. They need an open and accessible port, and a known vulnerability. If you remember the "Iptables Weekly Log Summary" in the opening section above, many of those may have all been the result of this type of scan. If you've followed the steps in this HOWTO, you should be reasonably safe here. This one is easy enough to deflect. 8.4.4. Script KiddiesA "script kiddie" is a "cracker" wanna be who doesn't know enough to come up with his/her own exploits, but instead relies on "scripts" and exploits that have been developed by others. Like "worms", they are looking for easy victims, and may similarly scan large address ranges looking for specific ports with known vulnerabilities. Often, the actual scanning is done from already comprised systems so that it is difficult to trace it back to them. The script kiddie has a bag of ready made tricks at his disposal, including an arsenal of "rootkits" for various Operating Systems. Finding susceptible victims is not so hard, given enough time and address space to probe. The motives are a mixed bag as well. Simple mischief, defacement of web sites, stolen credit card numbers, and the latest craze, "Denial of Service" attacks (see below). They collect zombies like trophies and use them to carry out whatever their objective is. Again, the key here is that they are following a "script", and looking for easy prey. Like the worm threat above, a functional firewall and a few very basic precautions, should be sufficient to deflect any threat here. By now, you should be relatively safe from this nuisance. 8.4.5. Spoofed IPsHow easy is it to spoof an IP address? With the right tools, very easy. How much of a threat is this? Not much, for most of us, and is over-hyped as a threat. Because of the way TCP/IP works, each packet must carry both the source and destination IP addresses. Any return traffic is based on this information. So a spoofed IP can never return any useful information to an attacker who is sending out spoofed packets. The traffic would go back to wherever that spoofed IP address was pointed. The attacker gets nothing back at all. This does have potential for "DoS" attacks (see below) where learning something about the targeted system is not important. And may be used for some general mischief making as well. 8.4.6. Targeted AttacksThe worm and wide ranging address type scans, are impersonal. They are just looking for any vulnerable system. It makes no difference whether it is a top secret government facility, or your mother's Window's box. But there are "black hats" that will spend a great deal of effort to get into a system or network. We'll call these "targeted" attacks since there has been a deliberate decision made to break in to a specific system or network. In this case, the attacker will look the system over for weaknesses. And possibly make many different kinds of attempts, until he finds a crack to wiggle through. Or gives up. This is more difficult to defend against. The attacker is armed and dangerous, so to speak, and is stalking his prey. Again, this scenario is very unlikely for a typical home system. There just generally isn't any incentive to take the time and effort when there are bigger fish to fry. For those who may be targets, the best defense here includes many of things we've discussed. Vigilance is probably more important than ever. Good logging practices and an IDS (Intrusion Detection System) should be in place. And subscribing to one or more security related mailing lists like BUGTRAQ. And of course, reading those alerts daily, and taking the appropriate actions, etc. 8.4.7. Denial of Service (DoS)"DoS" is another type of "attack" in which the intention is to disrupt or overwhelm the targeted system or network in such a way that it cannot function normally. DoS can take many forms. On the Internet, this often means overwhelming the victim's bandwidth or TCP/IP stack, by sending floods of packets and thus effectively disabling the connection. We are talking about many, many packets per second. Thousands in some cases. Or perhaps, the objective is to crash a server. This is much more likely to be targeted at organizations or high profile sites, than home users. And can be quite challenging to stop depending on the technique. And it generally requires the co-operation of networks between the source(s) and the target, so that the floods are stopped, or minimized, before they reach the targeted destination. Once they hit the destination, there is no good way to completely ignore them. "DDoS", Distributed Denial of Service, is where multiple sources are used to maximize the impact. Again, not likely to be directly targeted at home users. These are "slaves" that are "owned" by a cracker, or script kiddie, that are woken up and are targeted at the victim. There may be many computers involved in the attack. If you are home user, and with a dynamic IP address, you might find disconnecting, then re-connecting to get a new IP, an effective way out if you are the target. Maybe. 8.4.8. Brute Force"Brute force" attacks are where the attacker makes repetitive attempts at the same perceived weakness(es). Like a battering ram. A classic example would be where someone tries to access a telnet server simply by continually throwing passwords at it, hoping that one will eventually work. Or maybe crash the server. This doesn't require much imagination, and is not a commonly used tactic against home systems. By the way, this is one good argument against allowing remote root logins. The root account exists on all systems. It is probably the only one that this is true of. You'd like to make a potential attacker guess both the login name and password. But if root is allowed remote logins, then the attacker only needs to guess the password! 8.4.9. VirusesAnd now something not to worry about. Viruses seem to be primarily a Microsoft problem. For various reasons, viruses are not a significant threat to Linux users. This is not to say that it will always be this way, but the current virus explosion that plagues Microsoft systems, can not spread to Linux (or Unix) based systems. In fact, the various methods and practices that enable this phenomena, are not exploitable on Linux. So Anti-Virus software is not recommended as part of our arsenal. At least for the time being with Linux only networks. 8.5. LinksSome references for further reading are listed below. Not listed is your distribution's site, security page or ftp download site. You will have to find these on your own. Then you should bookmark them!
8.6. Editing Text FilesBy Bill Staehle All the world is a file. There are a great many types of files, but I'm going to stretch it here, and class them into two really broad families:
Binary files are meant to be read by machines, text files can be easily edited, and are generally read by people. But text files can be (and frequently are) read by machines. Examples of this would be configuration files, and scripts. There are a number of different text editors available in *nix. A few are found on every system. That would be '/bin/ed' and '/bin/vi'. 'vi' is almost always a clone such as 'vim' due to license problems. The problem with 'vi' and 'ed' is that they are terribly user unfriendly. Another common editor that is not always installed by default is 'emacs'. It has a lot more features and capability, and is not easy to learn either. As to 'user friendly' editors, 'mcedit' and 'pico' are good choices to start with. These are often much easier for those new to *nix. The first things to learn are how to exit an editing session, how to save changes to the file, and then how to avoid breaking long lines that should not be broken (wrapped). The 'vi' editor 'vi' is one of the most common text editors in the Unix world, and it's nearly always found on any *nix system. Actually, due to license problems, the '/bin/vi' on a Linux system is always a 'clone', such as 'elvis', 'nvi', or 'vim' (there are others). These clones can act exactly like the original 'vi', but usually have additional features that make it slightly less impossible to use. So, if it's so terrible, why learn about it? Two reasons. First, as noted, it's almost guaranteed to be installed, and other (more user friendly) editors may not be installed by default. Second, many of the 'commands' work in other applications (such as the pager 'less' which is also used to view man pages). In 'less', accidentally pressing the 'v' key starts 'vi' in most installations. 'vi' has two modes. The first is 'command mode', and keystrokes are interpreted as commands. The other mode is 'insert' mode, where nearly all keystrokes are interpreted as text to be inserted. ==> Emergency exit from 'vi' 1. press the <esc> key up to three times, until the computer beeps, or the screen flashes. 2. press the keys :q! <Enter> That is: colon, the letter Q, and then the exclamation point, followed by the Enter key. 'vi' commands are as follows. All of these are in 'command' mode:
NOTE: The four 'arrow' keys almost always work in 'command' or 'insert' mode. The 'ed' editor. The 'ed' editor is a line editor. Other than the fact that it is virtually guaranteed to be on any *nix computer, it has no socially redeeming features, although some applications may need it. A _lot_ of things have been offered to replace this 'thing' from 1975. ==> Emergency exit from 'ed' 1. type a period on a line by itself, and press <Enter> This gets you to the command mode or prints a line of text if you were in command mode. 2. type q and press <Enter>. If there were no changes to the file, this action quits ed. If you then see a '?' this means that the file had changed, and 'ed' is asking if you want to save the changes. Press q and <Enter> a second time to confirm that you want out. The 'pico' editor. 'pico' is a part of the Pine mail/news package from the University of Washington (state, USA). It is a very friendly editor, with one minor failing. It silently inserts a line feed character and wraps the line when it exceeds (generally) 74 characters. While this is fine while creating mail, news articles, and text notes, it is often fatal when editing system files. The solution to this problem is simple. Call the program with the -w option, like this: pico -w file_2_edit Pico is so user friendly, no further instructions are needed. It _should_ be obvious (look at the bottom of the screen for commands). There is an extensive help function. Pico is available with nearly all distributions, although it _may_ not be installed by default. ==> Emergency exit from 'pico' Press and hold the <Ctrl> key, and press the letter x. If no changes had been made to the file, this will quit pico. If changes had been made, it will ask if you want to save the changes. Pressing n will then exit. The 'mcedit' editor. 'mcedit' is part of the Midnight Commander shell program, a full featured visual shell for Unix-like systems. It can be accessed directly from the command line ( mcedit file_2_edit ) or as part of 'mc' (use the arrow keys to highlight the file to be edited, then press the F4 key). mcedit is probably the most intuitive editor available, and comes with extensive help. "commands" are accessed through the F* keys. Midnight Commander is available with nearly all distributions, although it _may_ not be installed by default. ==> Emergency exit from 'mcedit' Press the F10 key. If no changes have been made to the file, this will quit mcedit. If changes had been made, it will ask if you want to Cancel this action. Pressing n will then exit. 8.7. nmapLet's look at a few quick examples of what nmap scans look like. The intent here is to show how to use nmap to verify our firewalling, and system integrity. nmap has other uses that we don't need to get into. Do NOT use nmap on systems other than your own, unless you have permission from the owner, and you know it is not a violation of anyone's Terms of Service. This kind of thing will be taken as hostile by most people. As mentioned previously, nmap is a sophisticated port scanning tool. It tries to see if a host is "there", and what ports might be open. Barring that, what states those ports might be in. nmap has a complex command line and can do many types of "scans". See the man page for all the nitty gritty. A couple of words of warning first. If using portsentry, turn it off. It will drop the route to wherever the scan is coming from. You might want to turn off any logging also, or at least be aware that you might get copious logs if doing multiple scans. A simple, default scan of "localhost":
If you've read most of this document, you should be familiar with these services by now. These are some of the same ports we've seen in other examples. Some things to note on this scan: it only did 1500+ "interesting" ports -- not all ports. This can be configured differently if more is desirable (see man page). It only did TCP ports too. Again, configurable. It only picks up "listening" services, unlike netstat that shows all open ports -- listening or otherwise. Note the last "open" port here is 3000 is identified as "PPP". Wrong! That is just an educated guess by nmap based on what is contained in /etc/services for this port number. Actually in this case it is ntop (a network traffic monitor). Take the service names with a grain of salt. There is no way for nmap to really know what is on that port. Matching port numbers with service names can at times be risky. Many do have standard ports, but there is nothing to say they have to use the commonly associated port numbers. Notice that in all our netstat examples, we had two classes of open ports: listening servers, and then established connections that we initiated to other remote hosts (e.g. a web server somewhere). nmap only sees the first group -- the listening servers! The other ports connecting us to remote servers are not visible, and thus not vulnerable. These ports are "private" to that single connection, and will be closed when the connection is terminated. So we have open and closed ports here. Simple enough, and gives a pretty good idea what is running on bigcat -- but not necessarily what we look like to the outside world since this was done from localhost, and wouldn't reflect any firewalling or other access control mechanisms. Let's do a little more intensive scan. Let's check all ports -- TCP and UDP.
This is more than just "interesting" ports -- it is everything. We picked up a couple of new ones in the process too. We've seen these before with netstat, so we know what they are. That is the Junkbuster web proxy on port 8000/tcp and named on 32768/udp. This scan takes much, much longer, but it is the only way to see all ports. So now we have a pretty good idea of what is open on bigcat. Since we are scanning localhost from localhost, everything should be visible. We still don't know how the outside world sees us though. Now I'll ssh to another host on the same LAN, and try again.
I confess to tampering with the iptables rules here to make a point. Only two visible ports on this scan. Everything else is "closed". So says nmap. Once again:
Oops, I blocked ICMP (ping) while I was at it this time. One more time:
That's it. Notice how long that took. Notice ports are now "filtered" instead of "closed". How does nmap know that? Well for one, "closed" means bigcat sent a packet back saying "nothing running here", i.e. port is closed. In this last example, the iptables rules were changed to not allow ICMP (ping), and to "DROP" all incoming packets. In other words, no response at all. A subtle difference since nmap seems to still know there was a host there, even though no response was given. One lesson here, is if you want to slow a scanner down, "DROP" (or "DENY") the packets. This forces a TCP time out for the remote end on each port probe. Anyway, if your scans look like this, that is probably as well as can be expected, and your firewall is doing its job. A brief note on UDP: nmap can not accurately determine the status of these ports if they are "filtered". You probably will get a false-positive "open" condition. This has to do with UDP being a connectionless protocol. If nmap gets no answer (e.g. due to a "DROP"), it assumes the packets reached the target, and thus the port will be reported as "open". This is "normal" for nmap. We can play with firewall rules in a LAN set up to try to simulate how the outside world sees us, and if we are smart, and know what we are doing, and don't have a brain fart, we probably will have a pretty good picture. But it is still best to try to find a way to do it from outside if possible. Again, make sure you are not violating any ISP rules of conduct. Do you have a friend on the same ISP? 8.8. Sysctl OptionsThe "sysctl" options are kernel parameters that can be configured via the /proc filesystem. These can be dynamically adjusted at run-time. Typically these options are off if set to "0", and on if set to "1". Some of these have security implications, and thus is why we are here ;-) We'll just list the ones we think are relevant. Feel free to cut and paste these into a firewall script, or other file that is run during boot (like /etc/rc.local). Or your distribution may have their own way of tuning this. You can read up on what these mean in /usr/src/linux/Documentation/sysctl/README and other files in the kernel Documentation directories.
8.9. Secure AlternativesThis section will give a brief run down on secure alternatives to potentially insecure methods. This will be a hodge podge of clients and servers.
8.10. Ipchains and Iptables ReduxThis section offers a little more advanced look at some of things that ipchains and iptables can do. These are basically the same scripts as in Step 3 above, just with some more advanced configuration options added. These will provide "masquerading", "port forwarding", allow access to some user definable services, and a few other things. Read the comments for explanations. 8.10.1. ipchains II
8.10.2. iptables II
8.10.3. SummaryA quick run down of the some highlights... We added some host based access control rules: "blacklisted", and "trusted". We then showed several types of service and port based access rules. For instance, we allowed some very restrictive access to bigcat's POP3 server so we could connect only from our workplace. We allowed a very narrow rule for the ISP's DHCP server. This rule only allows one port on one outside IP address to connect to only one of our ports and only via the UDP protocol. This is a very specific rule! We are being specific since there is no reason to allow any other traffic to these ports or from these addresses. Remember our goal is the minimum amount of traffic necessary for our particular situation. So we made those few exceptions mentioned above, and all other services running on bigcat should be effectively blocked completely from outside connections. These are still happily running on bigcat, but are now safe and sound behind our packet filtering firewall. You probably have other services that fall in this category as well. We also have a small, home network in the above example. We did not take any steps to block that traffic. So the LAN has access to all services running on bigcat. And it is further "masqueraded", so that it has Internet access (different HOWTO), by manipulating the "forward" chain. And the LAN is still protected by our firewall since it sits behind the firewall. We also didn't impose any restrictive rules on the traffic leaving bigcat. In some situations, this might be a good idea. Of course, this is just a hypothetical example. Your individual situation is surely different, and would require some changes and likely some additions to the rules above. For instance, if your ISP does not use DHCP (most do not), then that rule would make no sense. PPP works differently and such rules are not needed. Please don't interpret that running any server as we did in this example is necessarily a "safe" thing to do. We shouldn't do it this way unless a) we really need to and b) we are running the current, safe version, and c) we are able to keep abreast of security related issues that might effect these services. Vigilance and caution are part of our responsibilities here too. 8.10.4. iptables mini-meJust to demonstrate how succinctly iptables can be configured in a minimalist situation, the below is from the Netfilter team's Rusty's Really Quick Guide To Packet Filtering:
This simple script will allow all outbound connections that we initiate, i.e. any NEW connections (since the default policy of ACCEPT is not changed). Then any connections that are "ESTABLISHED" and "RELATED" to these are also allowed. And, any connections that are not incoming from our WAN side interface, ppp0, are also allowed. This would be lo or possibly a LAN interface like eth1. So we can do whatever we want, but no unwanted, incoming connection attempts are allowed from the Internet. None. This script also demonstrates the creation of a custom chain, defined here as "block", which is used both for the INPUT and FORWARD chains. |