Open agroce opened 5 years ago
It seems like a lot of fuzzers work similarly, and that a lot of the front-ends would end up looking almost the same. Would it be a good idea to implement a base DeepStateFrontend
object in bin/deepstate/common.py
that does the work of consuming CLI args, and creating and calling the fuzzer command passed to subprocess.call
?
Not entirely sure how the whole front-end base API would work, but here would be a rough example for writing a front-end for Angora:
class Angora(DeepStateFrontend):
def __init__(self, fuzzer_envvar):
"""
the base object would determine path to
fuzzer binary based on envvar, and perform
other checks
"""
super().__init__(fuzzer_envvar)
def make_command(self, args_dict):
"""
maps dict() of deepstate-angora cli args
to cli command for fuzzer. Implementer can override
if arg flags differ from default
"""
super(Angora, self).make_command(args_dict)
def execute(self):
"""
takes fuzzer and command from make_command() and
executes with subprocess call
"""
super(Angora, self).execute()
One feature that I think this could also introduce would be the ability to interface compile-time instrumentation through the front-end (for the fuzzers that support it).
I'll have to think, but this sounds reasonable. Can you elaborate on what we might do with compile-time instrumentation here?
For example, maybe something like this:
$ deepstate-angora --compile harness.c
The front-end finds the path to angora-clang
, (or angora-clang++
for .cpp
) and compiles accordingly. A corresponding method in the base DeepStateFrontend
object could look like this:
class Angora(DeepStateFrontend):
...
def compile_and_instrument(self, test_file):
self.compiler_path = self.fuzzer_path + "/bin/angora_clang"
# ... subprocess.call to perform compilation and instrumentation
# with the appropriate flags
This goes back to the idea of using DeepState to interface a lot of the steps needed to perform fuzzing, so a user would just need to use a deepstate-{fuzzer_name}
front-end to do both compilation and the actual fuzz testing.
Getting it to work with more complex build stuff will be tricky, but for single files etc. yeah makes a lot of sense
Can we also include relevant functions to address #195? (for instance a function to save and load inputs for every fuzzer/symbolic execution engine)
Seems reasonable. a "grab current inputs" and "reseed" API
Then implement ensemble at a higher level of abstraction
Keep in mind that we need to do some additional research for "concolic execution", since I don't think Manticore has an API for that (but maybe angr has?)
Sure, though we can make a fuzzer ensemble without using any symex (other than what Eclipser and angora do)
We currently have a dedicated front-end to Eclipser (https://github.com/trailofbits/deepstate/blob/master/bin/deepstate/eclipser.py), which makes it easy to use Eclipser with DeepState. However, even using AFL (for example) requires a bit more typing and user suffering. In the case of AFL, it's not urgent, but there are likely other fuzzers (esp. file-based ones) where figuring out the right command line, or putting the tests generated in a useful location, is more difficult.