Skip to content

What afl-fuzz Is Bad At

American fuzzy lop is a polished and effective fuzzing tool. It has found tons of bugs and there are any number of blog posts talking about that. Here we’re going to take a quick look at what it isn’t good at. For example, here’s a program that’s trivial to crash by hand, that afl-fuzz isn’t likely to crash in an amount of time you’re prepared to wait:

#include <stdlib.h>
#include <stdio.h>

int main(void) {
  char input[32];
  if (fgets(input, 32, stdin)) {
    long n = strtol(input, 0, 10);
    printf("%ld\n", 3 / (n + 1000000));
  return 0;

The problem, of course, is that we’ve asked afl-fuzz to find a needle in a haystack and its built-in feedback mechanism does not help guide its search towards the needle. Actually there are two parts to the problem. First, something needs to recognize that divide-by-zero is a crash behavior that should be targeted. Second, the search must be guided towards inputs that result in a zero denominator. A finer-grained feedback mechanism, such as how far the denominator is from zero, would probably do the job here. Alternatively, we could switch to a different generation technology such as concolic testing where a solver is used to generate test inputs.

The real question is how to get the benefits of these techniques without making afl-fuzz into something that it isn’t — making it worse at things that it is already good at. To see why concolic testing might make things worse, consider that it spends a lot of time making solver calls instead of actually running tests: the base testing throughput is a small fraction of what you get from plain old fuzzing. A reasonable solution would be to divide up the available CPU time among strategies; for example, devote one core to concolic testing and seven to regular old afl-fuzz. Alternatively, we could wait for afl-fuzz to become stuck — not hitting any new coverage targets within the last hour, perhaps — and then switch to an alternate technique for a little while before returning to afl-fuzz. Obviously the various search strategies should share a pool of testcases so they can interact (afl-fuzz already has good support for communication among instances). Hopefully some concolic testing people will spend a bit of time bolting their tools onto afl-fuzz so we can play with these ideas.

A variant of the needle-in-a-haystack problem occurs when we have low-entropy input such as the C++ programs we would use to test a C++ compiler. Very few ascii strings are C++ programs and concolic testing is of very little help in generating interesting C++. The solution is to build more awareness of the structure of C++ into the testcase generator. Ideally, this would be done without requiring the somewhat elaborate effort we put into Csmith. Something like LangFuzz might be a good compromise. (Of course I am aware that afl-fuzz has been used against clang and has found plenty of ways to crash it! This is great but the bugs being found are not the kind of semantic bugs that Csmith was designed to find. Different testcase generators solve different problems.)

So we have this great fuzzing tool that’s easy to use, but it also commonly runs up against testing problems that it can’t solve. A reasonable way forward will be for afl-fuzz to provide the basic framework and then we can all build extensions that help it reach into domains that it couldn’t before. Perhaps Valgrind is a good analogy: it provides not only an awesome baseline tool but also a rich infrastructure for building new tools.

[Pascal Cuoq gave me some feedback on this post including a bunch of ideas about afl-fuzz-related things that he’ll hopefully blog about in the near future.]

UB Canaries

If you report an undefined behavior bug, a common reaction from software developers is “So what? Our code works just fine.” As a random example, here is a discussion I had with Rasmus Lerdorf about five years ago about some UBs in the PHP interpreter. One might point out that it wasn’t a very mature exchange but I wasn’t even 40 yet at the time. (Earlier I had an example here from OpenSSL but this one is more suitable.)

An idea I’ve kicked around for a long time, but only got around to implementing this week, is a collection of canaries for undefined behavior: little test programs that automate the process of determining whether a given compiler configuration is willing to exploit particular UBs. Here’s the UB-canary repo. The idea behind canaries is that since they’re easy to run, they provide a quick and easy way to figure out which UBs the users of a particular compiler should be especially worried about. The definition of “exploit” requires a bit of care: exploitation only makes sense in terms of an expectation held by a C/C++ programmer. For example, I might expect that when a negative number is shifted to the left, the result is the same as if the number had been cast to unsigned before being left-shifted. Of course we need to be completely clear with ourselves that any such expectation has nothing to do with the language standard and everything to do with what a particular compiler happens to do, either because the providers of that compiler are unwilling to exploit that UB or just because they have not gotten around to exploiting it yet. When no real guarantee from the compiler provider exists, we like to say that as-yet unexploited UBs are time bombs: they’re waiting to go off next month or next year when the compiler gets a bit more aggressive.

Below are some results from various versions of GCC and LLVM on Ubuntu Linux. Each compiler was tested at -O0, -O1, -O2, -Os, and -O3; the number in each part of the table indicates the number of optimization options at which the compiler was willing to exploit the UB. Most compilers were targeting x86 but for some of the older versions of GCC (4.0, 4.1, 4.2) I could no longer get that to work due to some sort of libc problem, so they were targeting x86-64. I doubt that affects the results much.

Here are the results for GCC:

4.0 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9
addr-null 5 5 5 5 5 5 0 0 0 0
array-oob 0 0 0 0 0 0 0 0 0 0
dangling-pointer 4 4 4 4 4 4 4 4 4 4
int-min-mod-minus-1 5 5 5 5 5 5 5 5 5 5
memcpy-overlap 5 5 5 5 5 5 5 5 5 5
modify-string-literal 5 5 5 5 5 5 5 5 5 5
pointer-casts 4 4 4 4 4 0 0 0 0 0
shift-by-bitwidth 2 2 2 3 3 3 3 3 4 4
signed-integer-overflow 2 5 3 3 3 3 3 3 3 3
signed-left-shift 0 0 0 0 0 0 0 0 0 0
strict-aliasing 3 3 3 0 2 0 3 3 3 3
uninitialized-variable 0 0 0 1 1 1 1 1 1 1

And for LLVM:

2.9 3.0 3.1 3.2 3.3 3.4 3.5 3.6
addr-null 0 0 0 0 0 0 0 0
array-oob 0 0 0 0 0 0 0 0
dangling-pointer 3 3 3 3 3 3 3 3
int-min-mod-minus-1 5 5 5 5 5 5 5 5
memcpy-overlap 5 5 5 5 5 5 5 5
modify-string-literal 5 5 5 5 5 5 5 5
pointer-casts 5 5 5 5 5 5 5 5
shift-by-bitwidth 3 3 3 3 3 3 3 3
signed-integer-overflow 4 4 4 4 4 4 4 4
signed-left-shift 0 0 0 0 0 0 0 0
strict-aliasing 3 3 3 3 3 3 3 3
uninitialized-variable 4 4 4 4 4 4 4 4

Click on the links to see the corresponding assumptions. It is interesting that GCC shows a lot more variance in UB exploitation than LLVM; I suspect that this is simply because the GCC versions shown here span a longer period of time than the LLVM versions (~10 years vs. ~4 years). It is also interesting that UB exploitation is not getting (as one might have feared) uniformly worse. Although overlapping memcpy() calls were always exploited by the compilers tested here, that is not the case on my Mac. I believe there have been compilers that exploited the signed-left-shift UBs but either my canaries aren’t good enough to find them or else those behaviors did not happen to be seen in the releases I choose here. Modifying a string literal never works on a modern Linux machine but it continues to work on embedded platforms that lack memory protection (I just saw some code doing this the other day). The INT_MIN % -1 case is interesting since it has a sensible answer and also my reading of the C99 standard is that this construct is not undefined; it is explicitly undefined in the newer standards.

Please let me know if:

  • You notice a bug in any UB canary or in the (admittedly crappy) driver script.
  • You know of a compiler that exploits an UB but where I don’t have a canary detecting the exploitation. A pull request would be ideal.

Canary code comes from a variety of sources and I have tried to give credit in the individual files. UB canaries are related to several ongoing UB-protection efforts but I want to in particular mention the Cerberus C Survey that is being run by Peter Sewell’s research group at Cambridge. Their goal is to create a mechanized semantics for C as it is really used, as opposed to the somewhat fictitious language described in the standards. If you are a C/C++ developer who cares about UB, please take a few minutes to fill out the survey.

Inexpensive CPU Monster

Rather than using the commercial cloud, my group tends to run day-to-day jobs on a tiny cluster of machines in my office and then to use Emulab when a serious amount of compute power is required. Recently I upgraded some nodes and thought I’d share the specs for the new machines on the off chance this will save others some time:

processor Intel Core i7-5820K $ 380
CPU cooler Cooler Master Hyper 212 EVO $ 35
mobo MSI X99S SLI Plus $ 180
RAM CORSAIR Vengeance LPX 16GB (4 x 4GB) DDR4 SDRAM 2400 $ 195
SSD SAMSUNG 850 Pro Series MZ-7KE256BW $ 140
video PNY Commercial Series VCG84DMS1D3SXPB-CG GeForce 8400 $ 49
case Corsair Carbide Series 200R $ 60
power supply Antec BP550 Plus 550W $ 60

These machines are well-balanced for the kind of work I do, obviously YMMV. The total cost is about $1100. These have been working very well with Ubuntu 14.04. They do a full build of LLVM in about 18 minutes, as opposed to 28 minutes for my previously-fastest machines that were based on the i7-4770. I’d be interested to hear where other research groups get their compute power — everything in the cloud? A mix of local and cloud resources? This is an area where I always feel like there’s plenty of room for improvement.

Booster Test

Ever since learning that the space shuttle booster motors were manufactured and tested at ATK in Promontory Utah — not too far from where I live — I wanted to see one of the tests. I didn’t manage to do that before the shuttle program was shut down, but today I got to see something better: a test of an SLS booster, which is about 25% more powerful than an STS booster and more than twice as powerful as one of the big F-1 engines from the Saturn V.

Here’s a close-up video. On the other hand, this one shows what the test was like from the viewing area, in particular the 8 seconds it took the noise to reach us. The sound was very impressive, with enough low-frequency power to make my clothing vibrate noticeably, but it was not anywhere close to painfully loud. The flame was, however, painfully bright to look at. The nozzle was being vectored around during the test (I hadn’t realized that the solid rockets participate in guidance) but that wasn’t easy to see from a distance.

NASA socials give some inside access to people like me (and you, if you live in the USA and want to sign up next time) who have no official connection to the space program. Yesterday we got to tour the plant where the boosters are made. It was great to learn about techniques for mixing, casting, and curing huge amounts of propellant without getting air bubbles or other imperfections into the mix and without endangering workers. The buildings in this part of ATK have escape slides from all levels and are surrounded by big earthworks to deflect potential explosions upwards. It was also really cool to see the hardware for hooking boosters to the main rocket, for vectoring nozzles, and things like that. Alas, we weren’t allowed to take pictures on the tour.

ATK’s rocket garden at sunrise:

And the main event:

Adandoned Mineshaft

Due to my 8-year-old’s obsession with Minecraft, the abandoned mineshafts found in the game are an everyday topic of discussion around the house. He is saving up to buy a pickaxe — no joke. Since we needed a day trip for the long weekend, I thought we’d visit some actual mines in the Silver Island Mountains near the Utah-Nevada border. We followed an old road most of the way up a minor mountain and found the remains of either a processing facility or living quarters, and several mine shafts higher up. One of these apparently goes all the way through the mountain but we didn’t do much more than poke our heads in, since it was difficult to gauge how dangerous it was. This area was mined from the 1870s through the 1930s. Since the pathway connecting the mine shafts with the highest vehicle access was too narrow and steep for vehicles, we were forced to assume that ore was transported downhill using some combination of human and animal power — the level of physical effort implied here seemed to be lost on the 8-year-old, who wants a pickaxe more than ever. We saw a pair of pronghorn antelope — but failed to get decent pictures.

Lamus Peak:

Built-up trail used to move ore downhill through rugged country:

Mine shaft:

Another one, with some nice copper ore right at the entrance:

And a lot of nothing in the distance:

Instruction Synthesis is Fun and Weird

Synthesis is sort of a hot topic in PL research lately. It basically means “implement a specification automatically.” Of course, at some level this isn’t very different from what compilers have been doing for ages, if we consider the source language program to be the specification, but when we’re doing synthesis the connotation is that the specification isn’t directly implementable using standard compiler techniques.

As of recently, we have an LLVM synthesizer for Souper. The basic method is from this excellent PLDI paper. In brief, you give the synthesis algorithm a specification and a library of components and it either gives you a way to wire together the components into a circuit that implements the specification or else it fails. Some other time I might explain the algorithm in detail, but here I wanted to give an example where we’re trying to figure out if a 64-bit word contains 16 repetitions of the same nibble value (this computation figured into many nibble sort submissions).

The specification in C is:

int nibs_same(unsigned long word) {
  unsigned long n00 = (word >> 60) & 0xf;
  unsigned long n01 = (word >> 56) & 0xf;
  unsigned long n02 = (word >> 52) & 0xf;
  unsigned long n03 = (word >> 48) & 0xf;
  unsigned long n04 = (word >> 44) & 0xf;
  unsigned long n05 = (word >> 40) & 0xf;
  unsigned long n06 = (word >> 36) & 0xf;
  unsigned long n07 = (word >> 32) & 0xf;
  unsigned long n08 = (word >> 28) & 0xf;
  unsigned long n09 = (word >> 24) & 0xf;
  unsigned long n10 = (word >> 20) & 0xf;
  unsigned long n11 = (word >> 16) & 0xf;
  unsigned long n12 = (word >> 12) & 0xf;
  unsigned long n13 = (word >>  8) & 0xf;
  unsigned long n14 = (word >>  4) & 0xf;
  unsigned long n15 = (word >>  0) & 0xf;
    (n00 == n01) & (n00 == n02) & (n00 == n03) & (n00 == n04) &
    (n00 == n05) & (n00 == n06) & (n00 == n07) & (n00 == n08) &
    (n00 == n09) & (n00 == n10) & (n00 == n11) & (n00 == n12) &
    (n00 == n13) & (n00 == n14) & (n00 == n15);

I won’t bother to give the LLVM version of this but it’s not too pretty, and obviously there’s not much that standard compiler optimization techniques can do to make it prettier. On the other hand, there’s a much better way to implement this specification:

int nibs_same(uint64_t word) {
  return word == rotate_right(word, 4);

LLVM doesn’t have a rotate instruction so the best we’ll do there is something like this:

define i32 @nibs_same(i64 %word) #0 {
  %1 = shl i64 %word, 60
  %2 = lshr i64 %word, 4
  %3 = or i64 %1, %2
  %4 = icmp eq i64 %3, %word
  %5 = zext i1 %4 to i32
  ret i32 %5

The question is: Will Souper generate this code from scratch? Turns out it does not, but on the other hand it generates this:

define i32 @nibs_same(i64 %word) #0 {
  %1 = lshr i64 %word, 28
  %2 = shl i64 %1, 32
  %3 = or i64 %1, %2
  %4 = icmp eq i64 %word, %3
  %5 = zext %4 to i32
  ret i32 %5

Or in C:

(((x >> 28) << 32) | (x >> 28)) == x

If you take a minute to look at this code, you’ll see that it is not implementing a rotate, but rather something odd and not obviously correct. It took me a while to convince myself that this is correct, but I believe that it is. The correctness argument uses these observations: (1) every nibble is compared to some other nibble and (2) it is irrelevant that the bitwise-or is presented with a pair of overlapping nibbles. Here’s the picture I drew to help:

                            >> 28
                            << 32

Also Alive can help:

  %1x = shl i64 %word, 60
  %2x = lshr i64 %word, 4
  %3x = or i64 %1x, %2x
  %4x = icmp eq i64 %3x, %word
  %5 = zext i1 %4x to i32
  %1 = lshr i64 %word, 28
  %2 = shl i64 %1, 32
  %3 = or i64 %1, %2
  %4 = icmp eq i64 %word, %3
  %5 = zext %4 to i32

Done: 1
Optimization is correct!

The drawback of Souper's code is that since it doesn't include a rotate operation that can be recognized by a backend, the resulting assembly won't be as good as it could have been. Of course if LLVM had a rotate instruction in the first place, Souper would have used it. Additionally, if we posit a superoptimizer-based backend for LLVM, it would be able to find the rotate operation that is well-hidden in Souper's synthesized code.

I find it kind of magical and wonderful that a SAT solver can solve synthesis problems like this. Souper's synthesizer is very alpha and I don't recommend using it yet, but the code is here if you want to take a look. Raimondas, my heroic postdoc, did the implementation work.

I'll finish up by noting that my first blog entry was posted five years ago today. Happy birthday, blog.

Update from Feb 10: A reader sent this solution which I hadn't seen before:

(x & 0x0FFFFFFFFFFFFFFF) == (x >> 4)

I think Souper missed it since we didn't include an "and" in its bag of components for this particular run.

Static Analysis Benchmarks

Many programmers would agree that static analysis is pretty awesome: it can find code defects that are very hard to find using testing and walkthroughs. On the other hand, some scientific validation of the effectiveness of static analysis would be useful. For example, this nice 2004 paper found that when five analyzers were turned loose against some C programs containing buffer overflows, only one of them was able to beat random guessing in a statistically significant fashion.

More recently, in 2014, some researchers at the Toyota InfoTechnology Center published a paper evaluating six static analyzers using a synthetic benchmark suite containing more than 400 pairs of C functions, where each pair contains one function exemplifying a defect that should be found using static analysis and the other function, while looking fairly similar, does not contain the defect. The idea is that a perfect tool will always find the bug and not find the not-bug, if you see what I mean. The results show that different tools have different strengths; for example, CodeSonar dominates in detection of concurrency bugs but PolySpace and QA-C do very well in finding numerical bugs. With a few exceptions, the tools do a good job in avoiding false positives. At this point I feel like I have to refer Coverity’s classic writeup about this kind of thing just because it’s awesome.

Shin’ichi Shiraishi, the lead author of the Toyota paper, contacted me asking if I’d help them release their benchmarks as open source, and I have done this: the repository is here, and here is a brief document describing the benchmarks. (I just want to make it clear that while I own the Github repo, the work here was entirely done by Shin’ichi and his colleagues — I added the simple autotooling, fixed a few portability problems in the benchmarks such as using intptr_t for integers that will be cast to and from pointer values, and did some other lightweight cleanup.) Anyway, I am very happy that Shin’ichi and his organization were able to release these benchmarks and we hope that they will be useful to the community.

Update from 2/25/15: Whoops — it looks like Coverity/Synopsys has a DeWitt clause in their EULA and based on this, they sent me a cease and desist notice. In accordance with their request, I have removed Shin’ichi’s group’s paper from the Github repo and also removed the link to that paper from this piece.

Nibble Sort Programming Contest

The problem is to sort the 4-bit pieces of a 64-bit word with (unsigned) smaller values towards the small end of the word. The nibble sort of 0xbadbeef is 0xfeedbba000000000. The function you implement will perform this sorting operation on a buffer of 1024 64-bit integers:

void nibble_sort(unsigned long *buf); 

I’ll give a small prize to the submitter of the entry that performs this operation the fastest on my Core i7-4770. If the winner makes non-trivial use of SIMD, I’ll give a prize to that entry and also to the fastest non-SIMD entry. I’ll use GCC 4.9.2 to compile your C99 code, which may use SIMD intrinsics and also inline assembly if you choose. There’s no requirement for portability. The machine runs Ubuntu 14.04 in 64-bit mode. If you require any command line options other than “gcc -std=gnu99 -O3″ you need to tell me that. You may assume that the buffer starts at the start of a cache line. The buffer will be loaded with uniformly chosen random values. Submit code by emailing me. Submit entries before the end of January 2015.

Here’s a slow (but hopefully correct) reference implementation takes about 180 380 microseconds to do the job:

int read_nibble(unsigned long w, int i) {
  assert(i >= 0 && i < 16);
  unsigned long res = w >> (i * 4);
  return res & 0xf;

void write_nibble(unsigned long *w, int i, int v) {
  assert(i >= 0 && i < 16);
  unsigned long mask = 0xf;
  mask <<= (i * 4);
  *w &= ~mask;
  unsigned long prom = v;
  prom <<= (i * 4);
  *w |= prom;

unsigned long nibble_sort_word(unsigned long arg) {
  for (int i = 0; i < 16; ++i) {
    int min = i;
    for (int j = i+1; j < 16; ++j) {
      if (read_nibble(arg, j) < read_nibble(arg, min))
        min = j;
    if (min != i) {
      int tmp = read_nibble(arg, i);
      write_nibble(&arg, i, read_nibble(arg, min));
      write_nibble(&arg, min, tmp);
  return arg;

void nibble_sort(unsigned long *buf) {
  for (int i=0; i<1024; i++)
    buf[i] = nibble_sort_word(buf[i]);

Update from Feb 1: Ok, contest is over. Thanks for the very impressive submissions, folks! Provisional results with and w/o SIMD are below. These are with turbo boost turned off. I’m now working on the blog post explaining the results, which is going to take some time because there are a lot of solutions here and some of them are very clever. I’m also still working to secure permission to use codes so I can put as many as possible in Github. There are fewer entries in the no-SIMD category because many codes had explicit use of vectors. Please let me know if you think I’ve made any mistakes.

Update from Feb 2: There were a few mistakes and omissions. Results that I hope are final are below.

regehr@regehr-M51AC:nibble_sort$ ./ 
              ns     ns / 8 bytes       entry name           errors
            1115             1.09       alexander4                0
            1238             1.21          arseny2                0
            1596             1.56          arseny1                0
            1755             1.71              uwe                0
            2630             2.57        pdewacht2                0
            4205             4.11         beekman2                0
            4466             4.36         beekman1                0
            4867             4.75       alexander3                0
            6084             5.94       alexander1                0
            6364             6.21         koorogi2                0
            8472             8.27             falk                0
            8520             8.32           jerome                0
           10331            10.09          vetter4                0
           10374            10.13           jepler                0
           10950            10.69       alexander2                0
           12096            11.81          vetter3               64
           12714            12.42         koorogi1                0
           14185            13.85              tom                0
           15532            15.17             rjk9                0
           16829            16.43          parnell                0
           16890            16.49        pdewacht1                0
           18804            18.36           chucky                0
           20041            19.57          burton2                0
           20649            20.17            nadav                0
           24908            24.32         bosmans1               60
           25104            24.52         bosmans2                0
           25486            24.89          vetter2                0
           28957            28.28             hans                0
           29928            29.23             anon                0
           30228            29.52            jrose                0
           31007            30.28           carlos                0
           32952            32.18            joris                0
           34562            33.75          vetter1                0
           45440            44.38             mats                0
           47511            46.40            frank                0
           50162            48.99            robin                0
           72762            71.06          grayson                0
           74465            72.72            rosen                0
           79752            77.88            payer                0
           92970            90.79          burton1                0
           94343            92.13           mentre                0
           95522            93.28             vlad                0
           97877            95.58           rogers                0
           99077            96.75             mike                0
          101913            99.52            bloom                0
          103945           101.51           jarkko                0
          109191           106.63            karim                0
          120928           118.09           justin                0
          160599           156.83           mikael                0
          200943           196.23           scotty                0
          416207           406.45              ref                0
regehr@regehr-M51AC:nibble_sort$ ./ NOSIMD
              ns     ns / 8 bytes       entry name           errors
            8593             8.39           jerome                0
           10321            10.08           jepler                0
           10346            10.10          vetter4                0
           10959            10.70       alexander2                0
           12105            11.82          vetter3               64
           15537            15.17             rjk9                0
           16842            16.45          parnell                0
           18847            18.41           chucky                0
           19995            19.53        pdewacht1                0
           21631            21.12            nadav                0
           25516            24.92          vetter2                0
           29418            28.73             hans                0
           29928            29.23             anon                0
           30529            29.81            jrose                0
           32956            32.18            joris                0
           34286            33.48             falk                0
           41028            40.07          burton2                0
           42308            41.32          vetter1                0
           44415            43.37         bosmans2                0
           45480            44.41             mats                0
           50179            49.00            robin                0
           53189            51.94            frank                0
           71285            69.61          grayson                0
           73883            72.15            rosen                0
           79744            77.88            payer                0
           94780            92.56          burton1                0
           94970            92.74           mentre                0
           95509            93.27             vlad                0
           98658            96.35           rogers                0
           98922            96.60             mike                0
          101722            99.34            bloom                0
          104880           102.42           jarkko                0
          113147           110.50            karim                0
          119168           116.38           justin                0
          415529           405.79              ref                0

Update from Feb 3: Here’s the Github repo containing all of the entries — let me know how it looks.

Buying Into Open Source Security

If you were given the opportunity to spend USD 100 million over five years to maximally improve the security of open source software, what would you do? Let’s just assume that the money comes with adequate administrative staff to manage awards and contracts so you can focus on technical issues. A few ideas:

  • Bug bounties, skewed towards remotely exploitable bugs and towards ubiquitous infrastructure such as OpenSSL and the Linux kernel. To get rid of USD 100 M in five years, we’ll probably need to make the bounties very large by current standards, or else give out a lot of them.
  • Contracts for compatible rewrites of crufty-but-important software in safe languages.
  • Contracts for aggressive cleanup and refactoring of things like OpenSSL.
  • Contracts for convincing demonstrations of the security of existing codes, in cases where it seems clear that rewrites are undesirable or impractical. These demonstrations might include formal verification, high-coverage test suites, and thorough code inspections.
  • Research grants and seed funding towards technologies such as unikernels, static analyzers, fuzzers, homomorphic encryption, Qubes/Bromium-kinda things, etc. (just listing some of my favorites here).
  • Contracts and grants for high-performance, open-source hardware platforms.

This post is motivated by the fact that there seems to be some under-investment in security-critical open source components like Bash and OpenSSL. I’ll admit that it is possible (and worrying) that USD 100 M isn’t enough to make much of a dent in our current and upcoming problems.

Testing with Pictures

Testing code is fun and hard and looking at the problem in different ways is always good. Here’s a picture representing the behavior of a saturating subtraction operation, where the horizontal axes represent the inputs and the output is vertical:

And here are some of the functions handed in by my students in the fall:

The last one represents one of the most common failure modes: failing to account for the asymmetry where INT_MIN has no additive inverse. The thing that I like about these images is how glaringly obvious the bugs are.

Hey, who knew Gnuplot could do animated gifs now? I guess old dogs can learn new tricks.

(Note: I posted some of these a few years ago, but I like the pictures so much I wanted to do it again.)