16.1 One Bug Can Ruin Your Whole Day . . .

The Unix security model makes a tremendous investment in the infallibility of the superuser and in the reliability of software that runs with the privileges of the superuser. If the superuser account is compromised, then the system is left wide open—hence, our many admonitions in this book to protect the superuser account and restrict the number of people who must know the superuser password.

Unfortunately, even if you prevent users from logging into the superuser account, many Unix programs need to run with some sort of administrative privileges. Many of these programs are set up to run with superuser privileges—typically by having them run as SUID root programs, by having the programs launched when the computer starts up, or by having them started by other programs running with superuser privileges (the common manner in which network servers are started). A single bug in any of these complicated programs can compromise the safety of your entire system. Furthermore, the environment and trusted inputs to these programs also need to be protected to prevent unexpected (and unwanted!) behavior.^[2] This characteristic is a security architecture design flaw, but it is basic to the design of Unix and is not likely to change.

^[2] Many of these programs should not run as superuser, but instead should run under another user that has a somewhat more restricted set of privileges.

16.1.1 The Lesson of the Internet Worm

One of the best examples of how a single line of code in a program can result in the compromise of thousands of machines dates back to the pre-dawn of the commercial Internet. The year was 1988, and a graduate student at Cornell University had discovered several significant security flaws in versions of Unix that were widely used on the Internet. Using his knowledge, the student created a program (known as a worm) that would find vulnerable computers, exploit one of these flaws, transfer a copy of itself to the compromised system, and then repeat the process. The program infected between 2,000 and 6,000 computers within hours of being released. While that does not seem like a lot of machines today, in 1988 it represented a substantial percentage of the academic and commercial mail servers on the Internet. The Internet was effectively shut down for two days following the worm's release.

Although the worm used several techniques for compromising systems, the most effective attack in its arsenal was a buffer overflow attack directed against the Unix fingerd daemon.

The original fingerd program contained these lines of code:

                char line[512];
...
                line[0] = '\0';
                gets(line);

Because the gets( ) function does not check the length of the line read, a program that supplied more than 512 bytes of valid data would overrun the memory allocated to the line[] array and, ultimately, corrupt the program's stack frame. The worm contained code that used the stack overflow to cause the fingerd program to execute a shell; because at the time it was standard practice to run fingerd as the superuser, this shell inherited superuser access to the server computer. fingerd didn't need to run as superuser, but it was spawned as a root process during the system startup and never switched to a different user ID.^[3] Because fingerd's standard input and standard output file descriptors were connected to the TCP socket, the remote process that caused the overflow was given complete, interactive control of the system.

^[3] This was common practice at the time. It predated the inetd and its mechanism of spawning servers with other user IDs. It also was a vulnerable paradigm that has led to countless other break-ins over the years, and it still poses a trap for the unwary.

The fix for the fingerd program was simple: replace the gets( ) function with the fgets( ) function. Whereas gets( ) takes one parameter, the buffer, the fgets( ) function takes three arguments: the buffer, the size of the buffer, and the file handle from which to fetch the data:

fgets(line,sizeof(line),stdin);

When the original fingerd program was written, it was common practice among many programers to use gets( ) instead of fgets( )—probably because using gets( ) required typing fewer characters each time. Nevertheless, because of the way that the C programming language and the Standard IO library were designed, any program that used gets( ) to fill a buffer on the stack potentially had—and still has—this vulnerability.

Although it seems like ancient history now, this story continues to illustrate many important lessons:

The worm demonstrated that a single flaw in a single innocuous Internet server could compromise the security of an entire system—and, indeed, an entire network.
Many of the administrators whose systems were compromised by the worm did not even know what the fingerd program did and had not made a conscious decision to have the service running. Likewise, many of the security flaws that have been discovered in the years since have been with software that was installed by default and not widely used.^[4]

^[4] This is not restricted to Unix—it has been common in the Windows family of systems, too.
Although the worm did not use its superuser access to intentionally damage programs or data on computers that it penetrated, the program did result in significant losses. Many of those losses were the result of lost time, lost productivity, and the loss of confidence in the compromised systems. There is no such thing as a "harmless break-in."
The worm showed that flaws in deployed software might lurk for years before being exploited by someone with the right tools and the wrong motives. Indeed, the flaw in the finger code had been unnoticed for more than six years, from the time of the first Berkeley Unix network software release until the day that the worm ran loose. This illustrates a fundamental lesson: because a hole has never been discovered in a program does not mean that no hole exists. The fact that a hole has not been exploited today does not guarantee that the hole will not be exploited tomorrow.

Interestingly enough, the fallible human component of secure programming is illustrated by the same example. Shortly after the problem with the gets( ) subroutine was exposed, the Berkeley programming group went through all of its code and eliminated every similar use of the gets( ) call in a network server. Most vendors did the same with their code. Several people, including Spafford in his paper analyzing the operations and effects of the worm, publicly warned that uses of other library calls that wrote to buffers without bounds checks also needed to be examined. These included calls to the sprintf( ) routine, and byte-copy routines such as strcpy ( ). However, those admonitions were not heeded.

In late 1995, as we were finishing the second edition of this book, a new security vulnerability in several versions of Unix was widely publicized. It was based on buffer overruns in the syslog library routine. An attacker could carefully craft an argument to a network daemon such that, when an attempt was made to log it using syslog, the message overran the buffer and compromised the system in a manner hauntingly similar to the fingerd problem. After seven years, a close cousin to the fingerd bug was discovered. What underlying library calls contribute to the problem? The sprintf( ) library call does, and so do byte-copy routines such as strcpy( ).

In the summer of 2002, as we were working on the third edition of this book, not one but four separate overflow vulnerabilities were found in the popular OpenSSL security library, based on effectively the same vulnerability. In use on more than a million Internet servers, this SSL library is the basis of the SSL offering used by the Apache web server and all Unix SSL-wrapped mail services.

While many Unix security bugs are the result of poor programming tools and methods, even more regrettable is the failure to learn from old mistakes, and the failure to redesign the underlying operating system or programming languages so that this broad class of attacks will no longer be effective.^[5]

^[5] Some efforts have been made to make Unix fundamentally more resistant to buffer overflow attacks. Modern BSD systems offer a nonexecutable stack—even if an attacker overflows a buffer into stack space, the code they insert cannot be executed. Solaris has made nonexecutable stack available since Version 2.6 (as a kernel option in /etc/system) and automatically enabled it for setuid files in Solaris 9. For Linux systems, the Openwall patches (http://www.openwall.com) can provide similar functionality. However, even with nonexecutable stacks, buffer overflows can be exploited to run arbitrary code, crash privileged code, and otherwise disrupt expected behavior.

16.1.2 An Empirical Study of the Reliability of Unix Utilities

In December 1990, the Communications of the ACM published an article by Miller, Fredrickson, and So entitled "An Empirical Study of the Reliability of Unix Utilities" (Volume 33, issue 12, pp. 32-44). The paper started almost as a joke: a researcher was logged into a Unix computer from home, and the programs he was running kept crashing because of line noise from a poor modem connection. Eventually, Barton Miller, a professor at the University of Wisconsin, decided to subject the Unix utility programs from a variety of different vendors to a selection of random inputs and monitor the results.^[6]

^[6] The Fuzz archive, including source code and additional papers—including the 1995 paper, "Fuzz Revisited: A Re-examination of the Reliability of UNIX Utilities and Sources," by Barton Miller et al.—can be found at http://www.cs.wisc.edu/~bart/fuzz/fuzz.html.

16.1.2.1 What he found

The results were discouraging. Between 25% and 33% of the Unix utilities could be crashed or hung by supplying them with unexpected inputs—sometimes input that was as simple as an end-of-file on the middle of an input line. On at least one occasion, crashing a program tickled an operating system bug and caused the entire computer to crash. Many times, programs would freeze for no apparent reason.

In 1995 a new team headed by Miller repeated the experiment, this time running a program called Fuzz on nine different Unix platforms. The team also tested Unix network servers, and a variety of X Window System applications (both clients and servers). Here are some of the highlights:

According to the 1995 paper, vendors were still shipping a distressingly buggy set of programs: "...the failure rate of utilities on the commercial versions of Unix that we tested (from Sun, IBM, SGI, DEC, and NeXT) ranged from 15-43%."
Unix vendors don't seem to be overly concerned about bugs in their programs: "Many of the bugs discovered (approximately 40%) and reported in 1990 are still present in their exact form in 1995. The 1990 study was widely published in at least two languages. The code was made freely available via anonymous FTP. The exact random data streams used in our testing were made freely available via FTP. The identification of failures that we found were also made freely available via FTP; these include code fragments with file and line numbers for the errant code. According to our records, over 2000 copies of the...tools and bug identifications were fetched from our FTP sites...It is difficult to understand why a vendor would not partake of a free and easy source of reliability improvements."
The two lowest failure rates in the study were the Free Software Foundation's GNU utilities (failure rate of 7%) and the utilities included with the freely distributed Linux version of the Unix operating system (failure rate 9%).^[7] Interestingly enough, the Free Software Foundation has strict coding rules that forbid the use of fixed-length buffers. (Miller et al. failed to note that many of the Linux utilities were repackaged GNU utilities.)

^[7] We don't believe that 7% is an acceptable failure rate, either.

There were a few bright points in the 1995 paper. Most notable was the fact that Miller's group was unable to crash any Unix network server. The group was also unable to crash any X Window System server.

On the other hand, the group discovered that many X clients will readily crash when fed random streams of data. Others will lock up—and in the process, freeze the X server until the programs are terminated.

In 2000, Professor Miller and Justin Forrester ran the Fuzz tests a third time, although this time exclusively against Windows NT. Their testing revealed that they could crash or hang 45% of all programs expecting user input. When they tried sending random Win32 messages to applications (something any user can accomplish), they disrupted 100% of all applications!

16.1.2.2 Where's the beef?

Many of the errors that Miller's group discovered resulted from common programming mistakes with the C programming language: programmers wrote clumsy or confusing code that did the wrong things; programmers neglected to check for array boundary conditions; and programmers assumed that their char variables were of type unsigned, when in fact they were signed.

While these errors can certainly cause programs to crash when they are fed random streams of data, these errors are exactly the kinds of problems that can be exploited by carefully crafted streams of data to achieve malicious results. Think back to the Internet worm: if tested by the Miller Fuzz program, the original fingerd program would have crashed. But when presented with the carefully crafted stream that was present in the worm, the program gave its attacker a root shell!

What is somewhat frightening about the study is that the tests employed by Miller's group are among the least comprehensive known to testers: random, black-box testing. Different patterns of input could possibly cause more programs to fail. Inputs made under different environmental circumstances could also lead to abnormal behavior. Other testing methods could expose these problems whereas random testing, by its very nature, would not.

Miller's group also found that use of several commercially available tools enabled them to discover errors and perform other tests, including discovery of buffer overruns and related memory errors. These tools were readily available; however, vendors were apparently not using them.^[8]

^[8] In the last decade, several of the firms making these tools went out of business or switched to selling other products. The reason? There were insufficient sales of software-testing tools to remain viable!

Why don't vendors care more about quality? Well, according to many of them, they do care, but quality does not sell. Writing good code and testing it carefully is not a quick or simple task. It requires extra effort and extra time. The extra time spent on ensuring quality will result in increased cost, and an increase in time-to-market. To date, few customers (possibly including you, gentle reader) have indicated a willingness to pay extra for better-quality software. Vendors have thus put their efforts into what customers are willing to buy, such as new features. Although we believe that most vendors could do a better job in this respect (and some could do a much better job), we must be fair and point the finger at the user population, too.

In some sense, any program you write might fare as well as vendor-supplied software. However, that isn't good enough if the program is running in a sensitive role and could potentially be abused. Therefore, you must practice good coding habits, and pay special attention to common trouble spots.