goblin VS fuzzcheck-rs

On the backend side, a lot has changed. The biggest change is that the dependency on Boost has been completely removed, and everything has been reworked accordingly. This should make packaging easier, as yabridge now no longer depends on any system library other than the basic libraries needed to interact with X11. Some parts of Boost have been replaced by other headers-only libraries, while other parts now simply use custom implementations. All of this is explained in more detail in the 'Packaging notes' section of the changelog. Yabridgectl also lost its dependency on winedump, at least in most cases. It now tries to parse plugin libraries directly using the goblin binary parsing library. This should also speed up the syncing process. I did, however, run into one plugin that this new parser couldn't handle. If that happens then winedump will still be used instead.

For now the minidump creation is a fairly faithful port of the Breakpad code, but like I said some of that code is really old, so there's probably cases where taking a step back and rethinking the approach based on new kernel or OS capabilities or, instead of recreating process snapshotting for each non-Windows, just have a really good parser for each OSes crash format that does a transform. Rust is a fantastic language for writing those kinds of parsers, so that would definitely be an interesting avenue to investigate, especially since in the Linux case a lot of groundwork has already been done by goblin.

I do security-related projects in Rust, and goblin has been my go-to crate for any type of binary parsing (ELF/PE/Mach-O).

Fuzzcheck is a structure-aware fuzzer for rust. "Fuzzing" means feeding large amounts of data into a program and checking for crashes (Fuzzcheck also checks to make sure that all the properties your program should uphold – e.g. a sorting algorithm applied to a list of n items should always return a list of n items – are upheld). Fuzzcheck is an "evolutionary" fuzzer – this means that it generates a set of random inputs, sees what percentage of the program is executed for each input, and keeps inputs which have high levels of percentage of program executed. It then "mutates" these inputs – whereas fuzzers such as AFL/Hongfuzz/etc mutate raw bytes in place (e.g. they swap bytes at different positions, or insert a random byte at a given position to generate inputs similar to the chosen "high coverage" inputs), Fuzzcheck works directly on the Rust types (so it might swap the order of two items in a vec, or randomly insert a new item). It's a really powerful tool for finding lots of bugs.

If you want help with Win support (issues/8) maybe post it here to get it added to TWIR.

I am working on a code coverage viewer for my fuzzer (fuzzcheck). I described what I've done so far in this issue and I am hoping to release the first version within two weeks.

The implications for my fuzzer, fuzzcheck, are huge! Compiling fuzz tests is a lot easier. There should be no more need to create a separate fuzz folder, fuzz tests can be regular #[test] functions, private implementation details can be fuzz-tested as well, rust-analyser works as expected, documentation can be easily generated, etc. I can also attach a human-readable coverage report to every test case :)

Since I graduated, I have had a lot more time to work on fuzzcheck. I am trying to flesh it out, test it, and document it for a new release. It has always felt a bit rushed/experimental and now I am hoping to make it into something solid. I have also played with an egui interface for it, to visualise the tested code coverage, understand how the fuzzer’s decisions are made, and also to interactively tweak the fuzzer’s behaviour. It's a lot of work but it's slowly all coming together! :)

fuzzcheck-rs is really cool. It combines property-testing with fuzzing, getting the nice, structured nature of the former, and the coverage-driven search of the latter, but it works by mutating the structure directly instead of going through a bit string. So if you have a binary tree, going from A(B, C) to A(C, B) can be a single mutation away if that makes sense in your use case, instead of being arbitrarily far away in the bitstring approach.