Test a fuzzer what fuzzing challenges it can solve.
The goal is to have testcases that AFL++ can all complete - once they are all implemented.
Currently the different solving mechanisms have to be set as defined in src/AFL-fuzz-redqueen.c
.
The comparisons to honggfuzz and libfuzzer are biased, as the testcases are how AFL-fuzz sees things.
The libfuzzer/
directories has libfuzzer's applicable own testcases, of which AFL++ is only able to solve 64% (vs 86% honggfuzz and 92% libfuzzer).
The fuzzers and necessary compilers must be in the path: AFL-clang-fast
, AFL-fuzz
, hfuzz-clang
, honggfuzz
, clang
.
./test.sh TARGET [SPECIFIC_TESTCASE]
TARGET
can be (currently) one of:
- AFL++ (llvm_mode with CMPLOG)
- AFL++-gcc (gcc_plugin with CMPLOG)
- AFL++-qemu
- AFL++-frida
- libAFL
- honggfuzz
- libfuzzer
- symsan (via
test-symsan.sh
, use it's docker container) - symcc + qemu (via
test-symcc.sh
andtest-symqemu.sh
, use their docker containers) - manticore (via
test-manticore.sh
) - tritondse (via
test-tritondse.sh
) - fuzzolic (via
test-fuzzolic.sh
+ docker.io/ercoppa/fuzzolic-runner-v1:ubuntu2004)
Note that manticore, tritondse, fuzzolic and symqemu success can depend on compile options.
test-u8
- several chained 8 bits checkstest-u16
- several chained 16 bits checkstest-u32
- several chained 32 bits checkstest-u64
- several chained 64 bits checkstest-u128
- several chained 128 bits checkstest-u32-cmp
- several chained 32 bit lesser/greater checkstext-extint
- llvm _ExtInt() tests (does not work with afl++-gcc)test-float
- several chained float checkstest-double
- several chained double checkstest-longdouble
- several chained long double checkstest-memcmp
- several chained memcmp checkstest-strcmp
- several chained strncasecmp checkstest-transform
- different transforming string checkstest-crc32
- several chained crc32 checks
The testcases are WIP - some might be further refined in the future and more added.
On success the time to solve is displayed.
On failure the generated corpus files are displayed (so you can see how many finds there were). There is also an output log called TESTCASE.log
.
All from current repository state (AFL++ is CMPLOG instrumented and afl-fuzz -l3AT -Z
).
Solve time: 120 seconds for AFL++*/libafl/honggfuzz/libfuzzer
The current default libafl fuzzer (fuzzbench variant) does not have a feature to stop when a crash is found, hence only OK or FAIL as results.
Symcc/SymQEMU, symsan, fuzzolic, TritonDSE and Manticore are not fuzzers but solvers, hence no time restriction.
SymQEMU currently has zero solves so it has been removed.
testcase | AFL++ | libAFL | symsan | symcc | manticore | tritondse | fuzzolic | AFL++-qemu/frida | honggfuzz-2.5 | libfuzzer-13 |
---|---|---|---|---|---|---|---|---|---|---|
test-crc32 | 0m1,735s | OK | OK | OK | OK | OK | OK | 0m14,609s | FAIL | 0m14,207s |
test-double | 0m26,823s | FAIL | FAIL | FAIL | FAIL | FAIL | FAIL | FAIL | FAIL | FAIL |
test-extint | 0m0,429s | OK | FAIL | FAIL | FAIL | FAIL | FAIL | FAIL | FAIL | FAIL |
test-float | 0m4,657s | FAIL | FAIL | FAIL | FAIL | FAIL | FAIL | FAIL | FAIL | FAIL |
test-longdouble | 0m1,031s | FAIL | FAIL | FAIL | FAIL | FAIL | FAIL | FAIL | FAIL | FAIL |
test-memcmp | 0m0,837s | OK | OK | OK | OK | OK | OK | 0m6,494s | 0m1,005s | 0m0,308s |
test-strcmp | 0m0,835s | OK | OK | FAIL | FAIL | OK | FAIL | 0m5,727s | 0m1,004s | 0m1,040s |
test-transform | 0m28,351s | FAIL | FAIL | FAIL | FAIL | FAIL | FAIL | FAIL | FAIL | 0m58,311s |
test-u128 | 0m0,682s | FAIL | FAIL | OK | FAIL | OK | OK | FAIL | FAIL | FAIL |
test-u16 | 0m1,252s | OK | OK | OK | OK | OK | OK | 0m8,132s | 0m1,005s | 0m3,741s |
test-u32 | 0m0,844s | OK | OK | OK | OK | OK | OK | 0m5,185s | 0m1,004s | 0m2,887s |
test-u32-cmp | 0m1,332s | OK | OK | OK | OK | OK | OK | 1m42,470s | 0m6,404s | 0m0,454s |
test-u64 | 0m0,655s | OK | OK | OK | OK | OK | OK | 0m3,844s | 0m1,005s | 0m5,465s |
test-u8 | 0m2,263s | OK | OK | OK | OK | OK | OK | 0m18,186s | 0m1,004s | 0m1,370s |
(*) some of these transform solvings are very expensive and hence disabled, but can be enabled at compile time in AFL++
AFL++ has the most solves, but due to the many solve attempts overall fuzzing performance is decreased, as can be seen at https://www.fuzzbench.com/reports/experimental/2021-01-12-aflpp/. Interpretation: the slowest solver is the best in real-world fuzzing.
Just add them and send a PR.