Passive fingerprinting uses most of the same techniques as the active fingerprinting performed by Nmap. The difference is that a passive system simply sniffs the network, opportunistically classifying hosts as it observes their traffic. This is more difficult than active fingerprinting, since you have to accept whatever communication happens rather than designing your own custom probes. It is a valuable technique, but doesn't belong in a fundamentally active tool such as Nmap. Fortunately, Michal Zalewski has written the excellent p0f passive OS fingerprinting tool. He also devised a couple of the current Nmap OS fingerprinting tests. Another option is SinFP by GomoR, which supports both active and passive fingerprinting.

TCP/IP fingerprinting works well for distinguishing different operating systems, but detecting different versions of the same operating system can be troublesome. The company must change their stack in some way we can differentiate. Fortunately, many OS vendors regularly update their systems to comply with the latest standards. But what about those who don't? Most of them at least get around to fixing exploitable stack bugs eventually. And those fixes are easy to detect remotely. First send the exploit payload, be it a land attack, teardrop, ping of death, SYN flood, or WinNuke. Send one attack at a time, then immediately try to contact the system again. If it is suddenly non-responsive, you have narrowed down the OS to versions which didn't ship with the fix.

	Warning
	If you use denial of service (DoS) exploits as part of your OS detection suite, remember to perform those tests last.

TCP implementations have significant leeway in exactly how long they wait before retransmitting packets. The proof-of-concept tools Ring and Cron-OS are available to exploit this. They send a SYN packet to an open port, then ignore the SYN/ACK they receive rather than acknowledging it with an ACK (to complete the connection) or a RST (to kill it). The target host will resend the SYN/ACK several more times, and these tools track every subsecond of the wait. While some information can indeed be gleaned from this technique, there are several reasons that I haven't incorporated the patch into Nmap:

I have enumerated these reasons here because they also apply to some other proposed OS detection methods. I would love to add new tests, but they must be quick and require few packets. Messing with host firewall is unacceptable. I try to avoid making full TCP connections for stack fingerprinting, though that is done for OS detection as part of the version scanning system.

IP fragmentation is a complex system and implementations are riddled with bugs and inconsistencies. Possible tests could examine how overlapping fragments are assembled or time the defragmentation timeouts. These tests are avoided for Nmap because many firewalls and other inline devices defragment traffic at gateways. Thus Nmap may end up fingerprinting the firewall rather than the true destination host. In addition, fragments are difficult to send on some operating systems. Linux 2.6 kernels have a tendency to queue the fragments you are trying to send and assemble them itself before transmission.

The target host OS can often be guessed simply by looking at the ports which are open. Microsoft Windows machines often have TCP ports 135 and 139 open. Windows 2000 and newer also listen on port 445. Meanwhile, a machine running services on port 22 (ssh) and 631 (Internet Printing Protocol) is likely running Unix.

While this heuristic is often useful, it just isn't reliable enough for Nmap. Combinations of ports can be obscured by firewall rules, and most mainstream protocols are available on multiple platforms. OpenSSH servers can be run on Windows, and the “Windows SMB” ports can be serviced by Samba running on a Unix machine. Port forwarding clouds the issue even further. A machine which appears to be running Microsoft IIS might be a Unix firewall simply forwarding port 80 to a Windows machine.

For these reasons, Nmap does not consider open port numbers during TCP/IP stack fingerprinting. However, Nmap can use version detection information (see Chapter 7, Service and Application Version Detection) to separately discover operating system and device type information. By keeping the OS detection results discovered by OS detection and version detection separate, Nmap can gracefully handle a Checkpoint firewall which uses TCP port forwarding to a Windows web server. The stack fingerprinting results should be “Checkpoint Firewall-1” while version detection should suggest that the OS is Windows. Keep in mind that only a small fraction of version detection signatures include OS and device type information—we can only populate these fields when the application divulges the information or when it only runs on one OS or device type.