Rust BOFs for Cobalt Strike

This took me like 4 days (+2 days for an update), but I got it working... rust core + alloc for Cobalt Strike BOFs.
This is very much a PoC, but I'd love to see others playing around with it and contributing.

Building

1. Install mingw
2. Install nightly rust with the x86_64-pc-windows-gnu and i686-pc-windows-gnu toolchains
3. Run cargo install cargo-make
4. Run cargo make
5. ????
6. BOFit

Make your own

Edit the entry function in rustbof/src/lib.rs. You can add new args by using the bof_pack function in the aggressor script, just don't change the first two because those are the relocations.

How the fk

I feel like I want to write a blog post about it at some point, but for now, here was the process:

• How do I compile object files from rust?
  • rustc has an --emit=obj flag that will just emit the object files into the deps folder of the target
• A BOF is a single object file, not many. Rust compiles each component into its own .o file. How do I combine them?
  • ld has a feature called "relocatable" (-i) which is used for incremental linking
• The Cobalt Strike loader throws throwing a NullPointerException on beacon_inline_execute. Why?
  • Some decompilation and bytecode debugging led me to find that the CS COFF parser was choking on some symbols in the object. Turns out those symbols were just included as debug (file) info.
  • In my investigation I found that the OBJExecutable and OBJParser classes in cobaltstrike.jar have main functions that take the path to an object file and print a bunch of useful information!
• How do I remove uneeded symbols in an object file?
  • Well, strip works! strip actually has a --strip-uneeded flag that strips everything not needed for relocations, like debug info!
• Now the Cobalt Strike BOF loader complains about some undefined symbols like rust_oom and __rust_alloc. At this point I wasn't using alloc at all, but it was still being compiled and then added to the BOF via ld -i. How can I get rid of these symbols?
  • At first I just removed the object file for alloc, since I wasn't using it. Easy.
  • I think at one point I discovered the --gc-sections flag for ld, which allows you to define a root symbol via the -u flag and then it gets rid of any symbols that aren't ever referenced. That also fixes it.
• OK. At this point we have something that loads up and does not crash. Let's try to import a function from KERNEL32. How do we make rust use __imp_ symbols?
  • We can define the link name via #[link_name = "__imp_KERNEL32$OutputDebugStringA]
• The issue with the above is that the __imp_ symbols are supposed to be pointers to the import table and not functions themselves, so rust thinks that the symbol is a single pointer to a function and not a double pointer to a function. How the heck do we convince rust that a function pointer is actually a function pointer pointer in a clean way?
  • Well, it turns out we can use unsafe rust to make a pointer into a double pointer, but that's only half of the solution because I don't want to have to say unsafe { make my function pointer a double pointer}(args) every time I want to make a call.
  • It is acutally possible to make rust import symbols via the __imp_ method, but I was only able to get it working on variables and not functions, so it was kind of useless.
• How can I automatically cast the pointer on call?
  • If you didn't know, rust calls deref when a type is called. So you can wrap the function pointer in a type and then implement core::ops::Deref for that type to cast the pointer to a double pointer on the fly. The result is a successful function call.
• So now we can import functions, but now what about an allocator? I want to be able to use strings and vectors. That's the nice part of rust core, alloc is easy to implement! The allocator just creates and manages a heap via NTDLL's Rtl*Heap functions. I defined a global allocator that must be initialized to be used. Now I'm getting those undefined symbols from alloc again such as __rust_alloc and rust_oom. Why?
  • It turns out when rust links a binary it creates those symbols and points them to either the system allocator or a custom allocator if one is defined. I need to define them manually.
• Now I'm getting an undefined symbol that looks like the name of my global allocator. What is happening?
  • BOFs don't use the BSS and the allocator was getting put in the BSS. I just explicitly told rust to put the ALLOCATOR inside the .data section via #[link_section = ".data"]. Fixed.
• Now that I can allocate memory, let's try using the format! macro in alloc to make an allocated String. It crashes! What gives?
  • This one took me a while to track down. The BOF loader does not relocate things in .data and .rdata, only .text. The COFF parser has a function at pe.OBJExecutable.getRelocations that creates a relocation structure for symbols in .text, but nothing else.
  • The crash was happening in an virtual function table in .rdata that was not getting updated.
  • Regardless, the BOF loader doesn't support type 1 (IMAGE_REL_AMD64_ADDR64), which is what rust was generating for .rdata.
• How can I get the relocations applied?
  • Write my own bootstrapper to manually apply the relocations before I do anything crazy like vtable references
  • Not that hard, but requires some coordination from the Cobalt Strike client side
  • I used Cobalt Strike's OBJExecutable class to parse the COFF and the Parser class to pack in the extra relocations. This is pretty much what CS does in getRelocations. The rust side gets the info via the BOF arguments and then applies the relocations.
• How do I know where the beacon is putting the other sections?
  • The BOF loader in beacon is going to choose some place to put the .data and .rdata sections, but we don't know where those are from our code.
  • I ended up adding importable symbols at the beginning of the .text, .rdata, and .data sections via modifying the linker script. If anyone has better ideas here I'd love to hear them.
• When those symbols were imported as extern static usize variables one of them generates an undefined refptr symbol. How do I stop it from doing that?
  • I ended up importing them as functions and then casting those to usize. A hack, but it gets the job done.
  • After fixups, vtable calls work!
• On to 32-bit. When you try to load it a bunch of syms are undefined, including a bunch of stuff starting with __. How can we resolve that???
  • 32-bit adds an underscore to a bunch of stuff for some reason I could look up but don't care.
  • I changed the import symbol to be _imp_ instead of __imp_
    • Doing this broke 64 bit so I added a cfg_attr to make the import name have the correct number of underscores depending on the target
  • I added a cfg_attr to the __section_start__ symbols from the linker to use a link name minus one underscore for x86 only.
  • The final symbol is the dreaded chkstk... which I've dealt with at length
• How can we get chkstk?
  • It is defined in NTDLL, so we can just import NTDLL!_chkstk and define a __chkstk ourselves that calls it

yeah and now both 32 and 64 bit BOFs work.
I haven't tried anything too incredibly fancy yet, but let me know if there are issues.

How the fk, part 2

Well it's 2024 now and this is broken on newest CS and rust.
The rust developers added a extern global static u8 variable called __rust_no_alloc_shim_is_unstable in this commit
Some of my friends and I were trying to figure out how to deal with it.

Ok, so my solution is not pretty. Basically, the compiler is putting a reference to that global static in it's own .rdata subsection called .rdata$.refptr.__rust_no_alloc_shim_is_unstable. Something to do with COMDAT or some shit. That section cannot be found by the BOF linker. So why not just combine all .rdata$.refptr.* sections into .rdata. That works fine, except now that variable cannot be accessed.

You might say to yourself... huh __rust_alloc should call my allocator and that should be that. Well now something actually calls __rust_alloc for you and right before it does it does a read of __rust_no_alloc_shim_is_unstable. This causes the BOF to crash, because it's actually trying to find that symbol via .refptr.__rust_no_alloc_shim_is_unstable, which is empty, as the linker is supposed to fill it out or something. The symbol must be defined somewhere too, and as far as I can tell it is not defined in alloc or in our object file.

So my solution is just to declare __rust_no_alloc_shim_is_unstable and write the address of it to .refptr.__rust_no_alloc_shim_is_unstable. To write to the address of the refptr symbol I just import it via link name (as a function so I don't generate more refptr symbols sigh). And then I write to that location when the allocator is initialized. And it actually works...

Sometime in between rust 1.61 and rust 1.75 the rust devs also made it so the #[global_allocator] attribute actually generates the __rust_{alloc|alloc_zeroed|dealloc|realloc} symbols, so we don't need to provide those anymore, BUT we do need to initialize the allocator somewhere now. I put that code in bofentry::entrypoint.

At least as of CS 4.9 the builtin COFF parser (pe.ObjExecutable) does relocations inside of .data, but still not .rdata. I suppose we could just merge .rdata into data and not have to do any relocations at all. A task for another time perhaps. I noticed this because all of my addresses in .data were about doubled... so I was applying the relocations when they were already applied.

It also seems like rustc now generates IMAGE_REL_AMD64_REL32 relocation types which are a bit annoying to understand. I referenced a few sources and came up with some code that seems to work.

So it works again. I also found a neat trick to stop LLVM from prepending underscores to every symbol. Just put \x01 right before the name. Thanks alexchrichton.

Versions

I don't know if this will work with future versions of CS or rust. Maybe.
I've been using rust 1.75.0 nightly (2024-02-07) and CS 4.9.

Known issues

TBD I guess.

TODO

• Make the bofhelper library more robust
• Docs, always <3
• Add some Result action with proper error types
• Allow you to specify the args in the bof macro and then have those args passed directly to your entry function. It would be cool to completely abstract away the BofData structure!

Thanks

• TrustedSec's COFF parser really helped with this - https://github.com/trustedsec/COFFLoader
• Delorie COFF structure pages - https://delorie.com/djgpp/doc/coff
• This post on Rust intermediates - https://medium.com/@squanderingtime/manually-linking-rust-binaries-to-support-out-of-tree-llvm-passes-8776b1d037a4
• This post on debugging java bytecode - https://www.crowdstrike.com/blog/native-java-bytecode-debugging-without-source-code/
• Sleep & CS beacon docs
• The rust compiler devs for adding a undefined global static u8 ref and ruining two days of my life