afl-fuzz_approach.md

The afl-fuzz approach

AFL++ is a brute-force fuzzer coupled with an exceedingly simple but rock-solid instrumentation-guided genetic algorithm. It uses a modified form of edge coverage to effortlessly pick up subtle, local-scale changes to program control flow.

Simplifying a bit, the overall algorithm can be summed up as:

1. Load user-supplied initial test cases into the queue.

2. Take the next input file from the queue.

3. Attempt to trim the test case to the smallest size that doesn't alter the measured behavior of the program.

4. Repeatedly mutate the file using a balanced and well-researched variety of traditional fuzzing strategies.

5. If any of the generated mutations resulted in a new state transition recorded by the instrumentation, add mutated output as a new entry in the queue.

6. Go to 2.

The discovered test cases are also periodically culled to eliminate ones that have been obsoleted by newer, higher-coverage finds; and undergo several other instrumentation-driven effort minimization steps.

As a side result of the fuzzing process, the tool creates a small, self-contained corpus of interesting test cases. These are extremely useful for seeding other, labor- or resource-intensive testing regimes - for example, for stress-testing browsers, office applications, graphics suites, or closed-source tools.

The fuzzer is thoroughly tested to deliver out-of-the-box performance far superior to blind fuzzing or coverage-only tools.

Understanding the status screen

This section provides an overview of the status screen - plus tips for troubleshooting any warnings and red text shown in the UI.

For the general instruction manual, see README.md.

A note about colors

The status screen and error messages use colors to keep things readable and attract your attention to the most important details. For example, red almost always means "consult this doc" :-)

Unfortunately, the UI will only render correctly if your terminal is using traditional un*x palette (white text on black background) or something close to that.

If you are using inverse video, you may want to change your settings, say:

• For GNOME Terminal, go to Edit > Profile preferences, select the "colors" tab, and from the list of built-in schemes, choose "white on black".
• For the MacOS X Terminal app, open a new window using the "Pro" scheme via the Shell > New Window menu (or make "Pro" your default).

Alternatively, if you really like your current colors, you can edit config.h to comment out USE_COLORS, then do make clean all.

We are not aware of any other simple way to make this work without causing other side effects - sorry about that.

With that out of the way, let's talk about what's actually on the screen...

The status bar

american fuzzy lop ++3.01a (default) [fast] {0}

The top line shows you which mode afl-fuzz is running in (normal: "american fuzzy lop", crash exploration mode: "peruvian rabbit mode") and the version of AFL++. Next to the version is the banner, which, if not set with -T by hand, will either show the binary name being fuzzed, or the -M/-S main/secondary name for parallel fuzzing. Second to last is the power schedule mode being run (default: fast). Finally, the last item is the CPU id.

Process timing

  +----------------------------------------------------+
  |        run time : 0 days, 8 hrs, 32 min, 43 sec    |
  |   last new find : 0 days, 0 hrs, 6 min, 40 sec     |
  | last uniq crash : none seen yet                    |
  |  last uniq hang : 0 days, 1 hrs, 24 min, 32 sec    |
  +----------------------------------------------------+

This section is fairly self-explanatory: it tells you how long the fuzzer has been running and how much time has elapsed since its most recent finds. This is broken down into "paths" (a shorthand for test cases that trigger new execution patterns), crashes, and hangs.

When it comes to timing: there is no hard rule, but most fuzzing jobs should be expected to run for days or weeks; in fact, for a moderately complex project, the first pass will probably take a day or so. Every now and then, some jobs will be allowed to run for months.

There's one important thing to watch out for: if the tool is not finding new paths within several minutes of starting, you're probably not invoking the target binary correctly and it never gets to parse the input files that are thrown at it; other possible explanations are that the default memory limit (-m) is too restrictive and the program exits after failing to allocate a buffer very early on; or that the input files are patently invalid and always fail a basic header check.

If there are no new paths showing up for a while, you will eventually see a big red warning in this section, too :-)

Overall results

  +-----------------------+
  |  cycles done : 0      |
  |  total paths : 2095   |
  | uniq crashes : 0      |
  |   uniq hangs : 19     |
  +-----------------------+

The first field in this section gives you the count of queue passes done so far

• that is, the number of times the fuzzer went over all the interesting test cases discovered so far, fuzzed them, and looped back to the very beginning. Every fuzzing session should be allowed to complete at least one cycle; and ideally, should run much longer than that.

As noted earlier, the first pass can take a day or longer, so sit back and relax.

To help make the call on when to hit Ctrl-C, the cycle counter is color-coded. It is shown in magenta during the first pass, progresses to yellow if new finds are still being made in subsequent rounds, then blue when that ends - and finally, turns green after the fuzzer hasn't been seeing any action for a longer while.

The remaining fields in this part of the screen should be pretty obvious: there's the number of test cases ("paths") discovered so far, and the number of unique faults. The test cases, crashes, and hangs can be explored in real-time by browsing the output directory, see #interpreting-output.

Cycle progress

  +-------------------------------------+
  |  now processing : 1296 (61.86%)     |
  | paths timed out : 0 (0.00%)         |
  +-------------------------------------+

This box tells you how far along the fuzzer is with the current queue cycle: it shows the ID of the test case it is currently working on, plus the number of inputs it decided to ditch because they were persistently timing out.

The "*" suffix sometimes shown in the first line means that the currently processed path is not "favored" (a property discussed later on).

Map coverage

  +--------------------------------------+
  |    map density : 10.15% / 29.07%     |
  | count coverage : 4.03 bits/tuple     |
  +--------------------------------------+

The section provides some trivia about the coverage observed by the instrumentation embedded in the target binary.

The first line in the box tells you how many branch tuples already were hit, in proportion to how much the bitmap can hold. The number on the left describes the current input; the one on the right is the value for the entire input corpus.

Be wary of extremes:

• Absolute numbers below 200 or so suggest one of three things: that the program is extremely simple; that it is not instrumented properly (e.g., due to being linked against a non-instrumented copy of the target library); or that it is bailing out prematurely on your input test cases. The fuzzer will try to mark this in pink, just to make you aware.
• Percentages over 70% may very rarely happen with very complex programs that make heavy use of template-generated code. Because high bitmap density makes it harder for the fuzzer to reliably discern new program states, we recommend recompiling the binary with AFL_INST_RATIO=10 or so and trying again (see env_variables.md). The fuzzer will flag high percentages in red. Chances are, you will never see that unless you're fuzzing extremely hairy software (say, v8, perl, ffmpeg).

The other line deals with the variability in tuple hit counts seen in the binary. In essence, if every taken branch is always taken a fixed number of times for all the inputs that were tried, this will read 1.00. As we manage to trigger other hit counts for every branch, the needle will start to move toward 8.00 (every bit in the 8-bit map hit), but will probably never reach that extreme.

Together, the values can be useful for comparing the coverage of several different fuzzing jobs that rely on the same instrumented binary.

Stage progress

  +-------------------------------------+
  |  now trying : interest 32/8         |
  | stage execs : 3996/34.4k (11.62%)   |
  | total execs : 27.4M                 |
  |  exec speed : 891.7/sec             |
  +-------------------------------------+

This part gives you an in-depth peek at what the fuzzer is actually doing right now. It tells you about the current stage, which can be any of:

• calibration - a pre-fuzzing stage where the execution path is examined to detect anomalies, establish baseline execution speed, and so on. Executed very briefly whenever a new find is being made.
• trim L/S - another pre-fuzzing stage where the test case is trimmed to the shortest form that still produces the same execution path. The length (L) and stepover (S) are chosen in general relationship to file size.
• bitflip L/S - deterministic bit flips. There are L bits toggled at any given time, walking the input file with S-bit increments. The current L/S variants are: 1/1, 2/1, 4/1, 8/8, 16/8, 32/8.
• arith L/8 - deterministic arithmetics. The fuzzer tries to subtract or add small integers to 8-, 16-, and 32-bit values. The stepover is always 8 bits.
• interest L/8 - deterministic value overwrite. The fuzzer has a list of known "interesting" 8-, 16-, and 32-bit values to try. The stepover is 8 bits.
• extras - deterministic injection of dictionary terms. This can be shown as "user" or "auto", depending on whether the fuzzer is using a user-supplied dictionary (-x) or an auto-created one. You will also see "over" or "insert", depending on whether the dictionary words overwrite existing data or are inserted by offsetting the remaining data to accommodate their length.
• havoc - a sort-of-fixed-length cycle with stacked random tweaks. The operations attempted during this stage include bit flips, overwrites with random and "interesting" integers, block deletion, block duplication, plus assorted dictionary-related operations (if a dictionary is supplied in the first place).
• splice - a last-resort strategy that kicks in after the first full queue cycle with no new paths. It is equivalent to 'havoc', except that it first splices together two random inputs from the queue at some arbitrarily selected midpoint.
• sync - a stage used only when -M or -S is set (see fuzzing_in_depth.md:3c) Using multiple cores). No real fuzzing is involved, but the tool scans the output from other fuzzers and imports test cases as necessary. The first time this is done, it may take several minutes or so.

The remaining fields should be fairly self-evident: there's the exec count progress indicator for the current stage, a global exec counter, and a benchmark for the current program execution speed. This may fluctuate from one test case to another, but the benchmark should be ideally over 500 execs/sec most of the time - and if it stays below 100, the job will probably take very long.

The fuzzer will explicitly warn you about slow targets, too. If this happens, see the best_practices.md#improving-speed for ideas on how to speed things up.

Findings in depth

  +--------------------------------------+
  | favored paths : 879 (41.96%)         |
  |  new edges on : 423 (20.19%)         |
  | total crashes : 0 (0 unique)         |
  |  total tmouts : 24 (19 unique)       |
  +--------------------------------------+

This gives you several metrics that are of interest mostly to complete nerds. The section includes the number of paths that the fuzzer likes the most based on a minimization algorithm baked into the code (these will get considerably more air time), and the number of test cases that actually resulted in better edge coverage (versus just pushing the branch hit counters up). There are also additional, more detailed counters for crashes and timeouts.

Note that the timeout counter is somewhat different from the hang counter; this one includes all test cases that exceeded the timeout, even if they did not exceed it by a margin sufficient to be classified as hangs.

Fuzzing strategy yields

  +-----------------------------------------------------+
  |   bit flips : 57/289k, 18/289k, 18/288k             |
  |  byte flips : 0/36.2k, 4/35.7k, 7/34.6k             |
  | arithmetics : 53/2.54M, 0/537k, 0/55.2k             |
  |  known ints : 8/322k, 12/1.32M, 10/1.70M            |
  |  dictionary : 9/52k, 1/53k, 1/24k                   |
  |havoc/splice : 1903/20.0M, 0/0                       |
  |py/custom/rq : unused, 53/2.54M, unused              |
  |    trim/eff : 20.31%/9201, 17.05%                   |
  +-----------------------------------------------------+

This is just another nerd-targeted section keeping track of how many paths were netted, in proportion to the number of execs attempted, for each of the fuzzing strategies discussed earlier on. This serves to convincingly validate assumptions about the usefulness of the various approaches taken by afl-fuzz.

The trim strategy stats in this section are a bit different than the rest. The first number in this line shows the ratio of bytes removed from the input files; the second one corresponds to the number of execs needed to achieve this goal. Finally, the third number shows the proportion of bytes that, although not possible to remove, were deemed to have no effect and were excluded from some of the more expensive deterministic fuzzing steps.

Note that when deterministic mutation mode is off (which is the default because it is not very efficient) the first five lines display "disabled (default, enable with -D)".

Only what is activated will have counter shown.

Path geometry

  +---------------------+
  |    levels : 5       |
  |   pending : 1570    |
  |  pend fav : 583     |
  | own finds : 0       |
  |  imported : 0       |
  | stability : 100.00% |
  +---------------------+

The first field in this section tracks the path depth reached through the guided fuzzing process. In essence: the initial test cases supplied by the user are considered "level 1". The test cases that can be derived from that through traditional fuzzing are considered "level 2"; the ones derived by using these as inputs to subsequent fuzzing rounds are "level 3"; and so forth. The maximum depth is therefore a rough proxy for how much value you're getting out of the instrumentation-guided approach taken by afl-fuzz.

The next field shows you the number of inputs that have not gone through any fuzzing yet. The same stat is also given for "favored" entries that the fuzzer really wants to get to in this queue cycle (the non-favored entries may have to wait a couple of cycles to get their chance).

Next is the number of new paths found during this fuzzing section and imported from other fuzzer instances when doing parallelized fuzzing; and the extent to which identical inputs appear to sometimes produce variable behavior in the tested binary.

That last bit is actually fairly interesting: it measures the consistency of observed traces. If a program always behaves the same for the same input data, it will earn a score of 100%. When the value is lower but still shown in purple, the fuzzing process is unlikely to be negatively affected. If it goes into red, you may be in trouble, since AFL++ will have difficulty discerning between meaningful and "phantom" effects of tweaking the input file.

Now, most targets will just get a 100% score, but when you see lower figures, there are several things to look at:

• The use of uninitialized memory in conjunction with some intrinsic sources of entropy in the tested binary. Harmless to AFL, but could be indicative of a security bug.
• Attempts to manipulate persistent resources, such as left over temporary files or shared memory objects. This is usually harmless, but you may want to double-check to make sure the program isn't bailing out prematurely. Running out of disk space, SHM handles, or other global resources can trigger this, too.
• Hitting some functionality that is actually designed to behave randomly. Generally harmless. For example, when fuzzing sqlite, an input like select random(); will trigger a variable execution path.
• Multiple threads executing at once in semi-random order. This is harmless when the 'stability' metric stays over 90% or so, but can become an issue if not. Here's what to try:
  • Use afl-clang-fast from instrumentation - it uses a thread-local tracking model that is less prone to concurrency issues,
  • See if the target can be compiled or run without threads. Common ./configure options include --without-threads, --disable-pthreads, or --disable-openmp.
  • Replace pthreads with GNU Pth (https://www.gnu.org/software/pth/), which allows you to use a deterministic scheduler.
• In persistent mode, minor drops in the "stability" metric can be normal, because not all the code behaves identically when re-entered; but major dips may signify that the code within __AFL_LOOP() is not behaving correctly on subsequent iterations (e.g., due to incomplete clean-up or reinitialization of the state) and that most of the fuzzing effort goes to waste.

The paths where variable behavior is detected are marked with a matching entry in the <out_dir>/queue/.state/variable_behavior/ directory, so you can look them up easily.

CPU load

  [cpu: 25%]

This tiny widget shows the apparent CPU utilization on the local system. It is calculated by taking the number of processes in the "runnable" state, and then comparing it to the number of logical cores on the system.

If the value is shown in green, you are using fewer CPU cores than available on your system and can probably parallelize to improve performance; for tips on how to do that, see fuzzing_in_depth.md:3c) Using multiple cores.

If the value is shown in red, your CPU is possibly oversubscribed, and running additional fuzzers may not give you any benefits.

Of course, this benchmark is very simplistic; it tells you how many processes are ready to run, but not how resource-hungry they may be. It also doesn't distinguish between physical cores, logical cores, and virtualized CPUs; the performance characteristics of each of these will differ quite a bit.

If you want a more accurate measurement, you can run the afl-gotcpu utility from the command line.

Interpreting output

See #understanding-the-status-screen for information on how to interpret the displayed stats and monitor the health of the process. Be sure to consult this file especially if any UI elements are highlighted in red.

The fuzzing process will continue until you press Ctrl-C. At a minimum, you want to allow the fuzzer to complete one queue cycle, which may take anywhere from a couple of hours to a week or so.

There are three subdirectories created within the output directory and updated in real-time:

• queue/ - test cases for every distinctive execution path, plus all the starting files given by the user. This is the synthesized corpus.

         Before using this corpus for any other purposes, you can shrink
         it to a smaller size using the afl-cmin tool. The tool will find
         a smaller subset of files offering equivalent edge coverage.
  

• crashes/ - unique test cases that cause the tested program to receive a fatal signal (e.g., SIGSEGV, SIGILL, SIGABRT). The entries are grouped by the received signal.

• hangs/ - unique test cases that cause the tested program to time out. The default time limit before something is classified as a hang is the larger of 1 second and the value of the -t parameter. The value can be fine-tuned by setting AFL_HANG_TMOUT, but this is rarely necessary.

Crashes and hangs are considered "unique" if the associated execution paths involve any state transitions not seen in previously-recorded faults. If a single bug can be reached in multiple ways, there will be some count inflation early in the process, but this should quickly taper off.

The file names for crashes and hangs are correlated with the parent, non-faulting queue entries. This should help with debugging.

Visualizing

If you have gnuplot installed, you can also generate some pretty graphs for any active fuzzing task using afl-plot. For an example of how this looks like, see https://lcamtuf.coredump.cx/afl/plot/.

You can also manually build and install afl-plot-ui, which is a helper utility for showing the graphs generated by afl-plot in a graphical window using GTK. You can build and install it as follows:

sudo apt install libgtk-3-0 libgtk-3-dev pkg-config
cd utils/plot_ui
make
cd ../../
sudo make install

To learn more about remote monitoring and metrics visualization with StatsD, see rpc_statsd.md.

Addendum: status and plot files

For unattended operation, some of the key status screen information can be also found in a machine-readable format in the fuzzer_stats file in the output directory. This includes:

• start_time - unix time indicating the start time of afl-fuzz
• last_update - unix time corresponding to the last update of this file
• run_time - run time in seconds to the last update of this file
• fuzzer_pid - PID of the fuzzer process
• cycles_done - queue cycles completed so far
• cycles_wo_finds - number of cycles without any new paths found
• execs_done - number of execve() calls attempted
• execs_per_sec - overall number of execs per second
• corpus_count - total number of entries in the queue
• corpus_favored - number of queue entries that are favored
• corpus_found - number of entries discovered through local fuzzing
• corpus_imported - number of entries imported from other instances
• max_depth - number of levels in the generated data set
• cur_item - currently processed entry number
• pending_favs - number of favored entries still waiting to be fuzzed
• pending_total - number of all entries waiting to be fuzzed
• corpus_variable - number of test cases showing variable behavior
• stability - percentage of bitmap bytes that behave consistently
• bitmap_cvg - percentage of edge coverage found in the map so far
• saved_crashes - number of unique crashes recorded
• saved_hangs - number of unique hangs encountered
• last_find - seconds since the last find was found
• last_crash - seconds since the last crash was found
• last_hang - seconds since the last hang was found
• execs_since_crash - execs since the last crash was found
• exec_timeout - the -t command line value
• slowest_exec_ms - real time of the slowest execution in ms
• peak_rss_mb - max rss usage reached during fuzzing in MB
• edges_found - how many edges have been found
• var_byte_count - how many edges are non-deterministic
• afl_banner - banner text (e.g., the target name)
• afl_version - the version of AFL++ used
• target_mode - default, persistent, qemu, unicorn, non-instrumented
• command_line - full command line used for the fuzzing session

Most of these map directly to the UI elements discussed earlier on.

On top of that, you can also find an entry called plot_data, containing a plottable history for most of these fields. If you have gnuplot installed, you can turn this into a nice progress report with the included afl-plot tool.

Addendum: automatically sending metrics with StatsD

In a CI environment or when running multiple fuzzers, it can be tedious to log into each of them or deploy scripts to read the fuzzer statistics. Using AFL_STATSD (and the other related environment variables AFL_STATSD_HOST, AFL_STATSD_PORT, AFL_STATSD_TAGS_FLAVOR) you can automatically send metrics to your favorite StatsD server. Depending on your StatsD server, you will be able to monitor, trigger alerts, or perform actions based on these metrics (e.g.: alert on slow exec/s for a new build, threshold of crashes, time since last crash > X, etc.).

The selected metrics are a subset of all the metrics found in the status and in the plot file. The list is the following: cycle_done, cycles_wo_finds, execs_done,execs_per_sec, corpus_count, corpus_favored, corpus_found, corpus_imported, max_depth, cur_item, pending_favs, pending_total, corpus_variable, saved_crashes, saved_hangs, total_crashes, slowest_exec_ms, edges_found, var_byte_count, havoc_expansion. Their definitions can be found in the addendum above.

When using multiple fuzzer instances with StatsD, it is strongly recommended to setup the flavor (AFL_STATSD_TAGS_FLAVOR) to match your StatsD server. This will allow you to see individual fuzzer performance, detect bad ones, see the progress of each strategy...