Format Strings - An Interview with Chris Evans

By Kurt Seifried, [email protected]

In a previous article I covered the basics of format string attacks. This time I've interviewed Chris Evans, whom I quoted in the last article. Without further ado, here is the interview. Read it - you will learn something. (I did.)


Q: It appears to me that these format strings have been present a very long time. A CERT advisory mentioned them being in WuFTPD since 1993. Do you think attackers have known about them and been using them? (This certainly would be a convenient explanation for many mysterious unsolved break-ins.)
A: This is a very interesting question. It depends what you mean by "attackers." I doubt this problem was widely known in the underground cracker community. When that is the case, the exploit usually leaks to the public. I can happily entertain that a few highly skilled individuals knew about this issue, though. Finally, we should be wary of attributing any unsolved break-ins to format string bugs. Even if a compromised site was running daemons containing format string bugs, there is still the potential for undiscovered security bugs which are not of a format string nature.

As we know, string/buffer handling has traditionally been very buggy. The most obvious example of buggy buffer handling is the classic buffer overflow.

The problem is that programmers tend to just treat a string or buffer as a "chunk of memory." This gives the programmer tedious chores such as:

  • When copying or appending strings, make sure the destination buffer is large enough.
  • If the destination buffer isn't large enough, do we truncate? Allocate a new bigger buffer? Error out?

All the calculations and checks involved in some of the above ways of coding are easy to get wrong. Also, code readability is reduced. It can be difficult to analyze the logic of a program because it is buried in buffer parsing and management.

The solution is break the association between string/buffer, and "chunk of memory." Instead of manipulating and copying things all over the place, we can manipulate an opaque object which just happens to internally manage a buffer.

Suddenly, instead of:

  char buf[BUFSIZE];
  strcpy(buf, data);
  strcat(buf, more_data);

we have

  struct safe_buf the_buf;

The former case is easy to get wrong. The latter case introduces an API which is very hard to use in an insecure manner. That is the key thing.

The safe_buf thing in fact boils down to little more than a C version of a classic C++ string or buffer class. You can happily push some string parsing into the safe_buf code too, e.g. string splitting, string substitutions, etc.

It is my understanding that the highly secure mail server "qmail" employs this technique for secure string handling (although I haven't looked at the source to confirm). Have you seen a string/buffer handling bug in qmail lately ;-)

Q: Like buffer overflows, bugs are due to programmers. Do you think programmers will ever break themselves of these habits/mistakes? Or are development tools the answer, like source code scanners such as ITS4?
A: Broadly, I think highly secure code requires two things.
  1. A high quality implementation
  2. A design resistant to flaws in the implementation

Unfortunately, even the best and most security-aware programmers are only human, and make mistakes. High(er) quality implementations can be achieved with a number of measures;

  • Using security-aware and good programmers to write security sensitive code.
  • Thorough code auditing of security sensitive code. Source scanners have their place, but they are no substitute for one, or preferably more humans reviewing code.
  • Throwing away coding styles and/or APIs which are error prone. (There is a very interesting example I could give here which illustrates that buffer overflows can be a non-problem if you change your coding style.)

Once we have taken steps to maximize the quality of code, there may unfortunately still be problems left. The recent "rpc.statd" exploit illustrates this. rpc.statd is a well audited piece of code, but a format string bug slipped through because at the times of the audits, no-one was looking for them. This is where a fault tolerant design comes in.

A fault tolerant design essentially minimizes the degree of privilege that bugs are able to give attackers. Obviously, one way this can be achieved under UNIX is by running as little code as possible as root. Other tools available to UNIX programmers sometimes include chroot()/jailing facilities and capabilities.

Unfortunately, most code isn't fault tolerant. A lot of daemons and services just permanently keep high levels of privilege. Not because they need it, but because it makes coding easier. It takes a non-trivial amount of effort to factor out parts of a program which genuinely _need_ privilege. Take something like the OpenSSH server. It has a good quality of implementation, largely because its security critical nature has inspired many audits. However, if a security hole is ever found in OpenSSH, its severity is likely to be a full machine compromise. That really need not be the case.

In summary, a way to ensure security despite the presence of the always fallible programmer, is through auditing and fault tolerant design requirements.

Q: I'm curious to know if you have seen SubDomain (from WireX). If so, what do you think of it?
A: I have heard of SubDomain, from the same people, I expect this might be the same thing as CoDomain ( gives some blurb).

I don't know all the gory details, but SubDomain is another application of the "principle of least privilege," giving much better system resilience in the presence of security holes you don't yet know about.

The SubDomain documentation, at the above URL, mentions enhanced file access control a lot. Hopefully, not just files will be protected, but also network access and process access. To generalize, the entire kernel API could use some decent protection. I like to visualize this protection as a "syscall firewall."

The existence of multiple projects with similar goals (SubDomain, Janus) shows a frustration that chroot() and, more recently, Linux kernel capabilities, are a step in the right direction, but simply not fine grained enough to apply the principle of least privilege properly.

Q: Do you think that like buffer overflows and format strings, there are more of these "fundamental" problems lying around, waiting to be found?
A: In the security world, we have to assume that yes, there are more undiscovered fundamental problems lying around. It would be foolish and short-sighted not to.

If we keep this threat in mind while writing code, it should help lead to fault tolerant solutions. The "remote-root" severity of the recent holes in WuFTPD, BSD-ftpd, rpc.statd and, probably, LPRng, simply would not have occurred in the presence of code with a good design from a security point of view.

Q: You mention fault tolerance several times. What do you think of things such as Linux's kernel capabilities, The Openwall kernel patch and StackGuard from WireX? Are they an answer, or do we need more?
A: I'll take these three things separately.

Linux kernel capabilities - these are a useful tool for reducing the impact of a vulnerability in a program or daemon. A superb example is a network time daemon. Why should it run with full privilege when the single privilege "change the system clock time" is required? Capabilities solve this nicely.

Unfortunately, for some other uses, they are not really fine grained enough. Two quick examples are

  • Listening on a privileged network port. With capabilities it's "none of them" or "all of them." It would be nice to nominate the ability to listen on a specific network port, like you mention below.
  • Ability to change the ownership on a file. Currently, this is either "you can't," or, "change any file to any owner." Better granularity would specify a list of allowable previous and new owners.

Openwall kernel patch - an interesting collection of patches. Some of these are geared towards making certain types of exploit harder or impossible, which is good. Classic security hardening. By far the most discussed component of this patchset, though, is the "non-exec stack" patch. This is essentially a defense against your traditional stack buffer overflow. Unfortunately, stack overflows are still exploitable with this patch in place. It just requires a different exploit. People concentrate on why this is bad, i.e. you are just as exploitable with or without this patch. However, another important facet is that your average script kiddie will be foiled by this patch, and move onto another target which is not as well protected (obfuscated, some might say). This will remain true for as long as the number of installations without this protection is relatively large. If this protection because commonplace, exploits would simply be released targeted at systems with the non-exec stack patch in place.

StackGuard - another solution to stack based buffer overflows, but one which will make some stack overflow bugs unexploitable. This is a very useful layer of security to add. If your extensive code audit were to miss a stack overflow (and this happens), then you may still not be exploitable.

The direction I'd like to see things take, is better OS support for applications to describe the precise set of privilege they need. If done properly, most applications would run with sufficiently low privilege that a compromise would be only a fraction as serious as one is today.

Things like StackGuard have a place too - namely targeting a common flaw and preventing it being a problem. In an ideal world, we would just nail the problem in the first place. However, this is not an ideal world, and some instances of "problem" get missed during auditing.

Q: I have seen other projects in progress. One that intrigued me was the ability to assign a port to a user/group (a la file ownership) so that, say, "dnssrvr" could use port 53 and not ever need to touch root. If you had a wish list of such items, what would the top three be?
A: Nice question :)
  1. Better OS support for jailing

    By "jailing," I mean placing strong restrictions on what a process is allowed to do. The only thing close to this in common UNIX use at the moment is chroot(). This essentially is restricting what files a compromised process can execute or open/read. Preventing arbitrary file execution is very powerful, because via suid-root binaries, it is a way of leveraging higher privilege.

    This needs to be extended to the ability to deny other system interactions such as making/receiving network connections, signals, using shared memory, allocating memory, creating processes, etc.

    Turning the problem on its head, doing anything should be denied to a daemon/server, apart from things it specifically needs to do. For some network servers you need little more than read()/write() on an existing network connection, plus the ability to open a small subset of the filesystem read-only. All other syscalls should be denied, which also has the pleasant effect of protecting from many potential kernel flaws which might be used to gain privilege.

    I haven't looked at the implementation yet, but "Janus" sounds very promising along these lines;

    For a final different perspective on better jailing, consider a cracker who has found a security hole in a FTP server, which they then use to be able to execute arbitrary code. Imagine their frustration at not being able to execute /usr/bin/suidperl to gain root privs. Not being able to connect out to other systems. Not being able to DoS the local machine.. etc..

  2. Reclassifying existing UNIX rights as capabilities

    If I login as "chris" to my local machine, I get dropped at what is termed an unprivileged shell. In fact, it is not unprivileged in the slightest. I will find myself able to set up network servers. Connect out to other machines on the Internet. Run privileged binaries.

    It would be interesting to extend the capabilities system to make capabilities of some things that current operating systems allow by default to normal users.

  3. Rewrite/re-architecting of existing software

    Maybe this doesn't count in the context of the question you asked :) However, most existing software fails to use the privilege limitation/jailing facilities available to it. If better jailing facilities do become standard, then software will start to look _really_ outdated from a security standpoint. So, my wish list finishes with overhauling (or replacing) a lot of security sensitive software.


Last updated on 3/30/2002

Copyright Kurt Seifried 2002