#gadget #x86-64 #rop #jop

bin+lib xgadget

Fast, parallel, cross-variant ROP/JOP gadget search for x86/x64 binaries

16 releases (10 breaking)

0.11.1 Nov 24, 2023
0.10.0 Nov 21, 2023
0.8.0 Jun 12, 2022
0.5.1 Jun 12, 2021
0.1.2 Jul 3, 2020

#251 in Development tools

MIT license

170KB
3.5K SLoC

xgadget

crates.io docs.rs GitHub Actions License: MIT

Fast, parallel, cross-{patch,compiler}-variant ROP/JOP gadget search for x86 (32-bit) and x64 (64-bit) binaries. Uses the iced-x86 disassembler library.

This crate can be used as a CLI binary (Windows/Linux/MacOS) or a library (7 well-known dependencies, all Rust).

Quickstart

Install the CLI tool and show its help menu:

cargo install xgadget --features cli-bin    # Build on host (pre-req: https://www.rust-lang.org/tools/install)
xgadget --help                              # List available command line options

How do ROP and JOP code reuse attacks work?

  • Return Oriented Programming (ROP) introduced code-reuse attacks, after hardware mitigations (aka NX, DEP) made code-injection less probable (no simultaneous WRITE and EXECUTE memory permissions). An attacker with stack control chains together short, existing sequences of assembly (aka "gadgets") - should a leak enable computing gadget addresses in the face of ASLR. When contiguous ROP gadget addresses are written to a corrupted stack, each gadget's ending ret instruction pops the next gadget's address into the CPU's instruction pointer. The result? Turing-complete control over a victim process.

rop model

ROP Attack Model (recreated from: Bletsch et. al.)

  • Jump Oriented Programming (JOP) is a newer code reuse method which, unlike ROP, doesn't rely on stack control. The attack bypasses hardware-assisted shadow-stack implementations (e.g. Intel CET's shadow stack), and is limited but not prevented by prototype-insensitive indirect target checks (e.g. Intel CET's IBT). JOP allows storing a table of gadget addresses in any READ/WRITE memory location. Instead of piggy-backing on call-return semantics to execute a gadget list, a "dispatch" gadget (e.g. add rax, 8; jmp [rax]) controls table indexing. Chaining happens if each gadget ends with a jmp back to the dispatcher (instead of a ret).

jop model

JOP Attack Model (recreated from: Bletsch et. al.)

About

xgadget is a tool for Return-Oriented Programming (ROP) and Jump-Oriented Programming (JOP) exploit development. It's a fast, multi-threaded alternative to awesome tools like ROPGadget, Ropper, and rp.

The goal is supporting practical usage while simultaneously exploring unique and experimental features. To the best of our knowledge, xgadget is the first gadget search tool to be:

  • Fast-register-sensitive: Filters gadgets by register usage behavior, not just matches for a given regex, without SMT solving (more powerful, but often impractical).

    • --reg-overwrite [<OPT_REG(S)>...] - control any reg (no args) or specific regs (args)

    • --reg-mem-write [<OPT_REG(S)>...] - write mem indexed via any reg (no args) or specific regs (args)

    • --reg-no-write [<OPT_REG(S)>...] - don't write any reg (no args) or specific regs (args)

    • --reg-read [<OPT_REG(S)>...] - read any regs (no args) or specific regs (args)

    • --reg-mem-read [<OPT_REG(S)>...] - read mem indexed via any reg (no args) or specific regs (args)

    • --reg-no-read [<OPT_REG(S)>...] - don't read any regs (no args) or specific regs (args)

  • JOP-efficient: JOP search uses instruction semantics - not hardcoded regex for individual encodings.

    • Optionally filter to JOP "dispatcher" gadgets with flag --dispatcher
  • Cross-variant: Finds gadgets that work across multiple variants of a binary (e.g. anti-diversification for different program or compiler versions). Two strategies:

  1. Full-match - Same instruction sequence, same program counter: gadget fully re-usable. Example:
    • Gadget: pop rdi; ret;
    • Address (in all binaries): 0xc748d

full match

Cross-variant Full Match

  1. Partial-match - Same instruction sequence, different program counter: gadget logic portable. Example:
    • Gadget: pop rdi; ret;
    • Address in bin_v1.1: 0xc748d
    • Address in bin_v1.2: 0xc9106

partial match

Cross-variant Partial Match

  • This is entirely optional, you're free to run this tool on a single binary.

Other features include:

  • Supports ELF32, ELF64, PE32, PE32+, Mach-O, and raw files
  • Parallel across available cores, whether searching a single binary or multiple variants
  • Currently 8086/x86/x64 only (uses a speed-optimized, arch-specific disassembler)

CLI Examples

Run xgadget --help to enumerate available options.

  • Example: Search /usr/bin/sudo for reliable ways to control rdi:
xgadget /usr/bin/sudo --reg-only --reg-overwrite rdi
  • Example: Search for ROP gadgets that control the value of rdi, never read rsi or rdx, and occur at addresses that don't contain bytes 0x32 or 0x0d:
xgadget /usr/bin/sudo --rop --reg-overwrite rdi --reg-no-read rsi rdx --bad-bytes 0x32 0x0d
  • Example: Search /usr/bin/sudo for "pop, pop, {jmp,call}" gadgets up to 10 instructions long, print results using AT&T syntax:
xgadget /usr/bin/sudo --jop --reg-pop --att --max-len 10
  • Example: Same as above, except using a regex filter to match "pop, pop, {jmp,call}" instruction strings (slower/less-accurate here, but regex enables flexible search in general):
xgadget /usr/bin/sudo --regex-filter "^(?:pop)(?:.*(?:pop))*.*(?:call|jmp)" --att --max-len 10
  • Example: Examine the exploit mitigations binaries sudo and lighttpd have been compiled with:
xgadget /usr/bin/sudo /usr/sbin/lighttpd --check-sec
  • Example: List imported and internal symbols for lighttpd:
xgadget /usr/sbin/lighttpd --symbols

API Usage

Find gadgets:

use xgadget::{Binary, SearchConfig};

let max_gadget_len = 5;

// Search single binary
let bin = &[Binary::from_path("/path/to/bin_v1").unwrap()];
let gadgets =
    xgadget::find_gadgets(bin, max_gadget_len, SearchConfig::default()).unwrap();
let stack_pivot_gadgets = xgadget::filter_stack_pivot(gadgets);

// Search for cross-variant gadgets, including partial matches
let search_config = SearchConfig::default() | SearchConfig::PART;
let bins = &[
    Binary::from_path("/path/to/bin_v1").unwrap(),
    Binary::from_path("/path/to/bin_v2").unwrap(),
];
let cross_gadgets =
    xgadget::find_gadgets(bins, max_gadget_len, search_config).unwrap();
let cross_reg_pop_gadgets = xgadget::filter_reg_pop_only(cross_gadgets);

Custom filters can be created using the GadgetAnalysis object and/or functions from the semantics module. How the above filter_stack_pivot function is implemented:

use rayon::prelude::*;
use iced_x86;
use xgadget::{Gadget, GadgetAnalysis};

/// Parallel filter to gadgets that write the stack pointer
pub fn filter_stack_pivot<'a, P>(gadgets: P) -> P
where
    P: IntoParallelIterator<Item = Gadget<'a>> + FromParallelIterator<Gadget<'a>>,
{
    gadgets
        .into_par_iter()
        .filter(|g| {
            let regs_overwritten = g.analysis().regs_overwritten(true);
            if regs_overwritten.contains(&iced_x86::Register::RSP)
                || regs_overwritten.contains(&iced_x86::Register::ESP)
                || regs_overwritten.contains(&iced_x86::Register::SP)
            {
                return true;
            }
            false
        })
        .collect()
}

Why No Chain Generation?

Tools that attempt to automate ROP/JOP chain generation require heavyweight analysis - typically symbolic execution of an intermediate representation. This works well for small binaries and CTF problems, but tends to be error-prone and difficult to scale for large, real-world programs. At present, xgadget has a different goal: enable an expert user to manually craft stable exploits by providing fast, accurate gadget discovery.

Yeah, but can it do 10 OS kernels under 10 seconds?! Repeatable Benchmark Harness

To build a Docker container and connect to it:

user@host$ git clone git@github.com:entropic-security/xgadget.git
user@host$ cd xgadget
user@host$ docker build -t xgadget_bench_container .
user@host$ docker run -it xgadget_bench_container
root@container:/xgadget#

The final build step runs ./benches/bench_setup_ubuntu.sh. This script downloads and builds 10 consecutive Linux kernels (versions 5.0.1 to 5.0.10 - with x86_64_defconfig). Grab a coffee, it can take a while.

Once it's done, run cargo bench to search all 10 kernels for common gadgets (among other benchmarks):

root@container:/xgadget# cargo bench

On an i7-9700K (8C/8T, 3.6GHz base, 4.9 GHz max) machine with gcc version 8.4.0: the average runtime, to process all ten 54MB kernels simultaneously with a max gadget length of 5 instructions and full-match search for all gadget types (ROP, JOP, and syscall gadgets), is only 6.3 seconds! Including partial matches as well takes just 7.9 seconds.

Fast Exploit Similarity Score (FESS)

The --fess flag uses cross-variant gadget matching as a metric of binary similarity. It's an experiment in anti-diversification for exploitation. To view similarity scores for kernel versions 5.0.1, 5.0.5, and 5.0.10 within the container:

root@container# cd ./benches/kernels/
root@container# xgadget vmlinux-5.0.1 vmlinux-5.0.5 vmlinux-5.0.10 --fess
TARGET 0 - [ name: 'vmlinux-5.0.1' | fmt-arch: ELF-X64 | entry: 0x00000001000000 | exec bytes/segments: 21,065,728/2 ]
TARGET 1 - [ name: 'vmlinux-5.0.5' | fmt-arch: ELF-X64 | entry: 0x00000001000000 | exec bytes/segments: 21,069,824/2 ]
TARGET 2 - [ name: 'vmlinux-5.0.10' | fmt-arch: ELF-X64 | entry: 0x00000001000000 | exec bytes/segments: 21,069,824/2 ]

┌─────────────┬──────────────────────┬──────────────────────┬───────────────────────┐
 Gadget Type │ vmlinux-5.0.1 (base)  vmlinux-5.0.5 (diff)  vmlinux-5.0.10 (diff) 
├─────────────┼──────────────────────┼──────────────────────┼───────────────────────┤
  ROP (full)               108,380 │        7,351 (6.78%)            556 (0.51%) 
├─────────────┼──────────────────────┼──────────────────────┼───────────────────────┤
  ROP (part)                     - │      80,783 (74.54%)        78,053 (72.02%) 
├─────────────┼──────────────────────┼──────────────────────┼───────────────────────┤
  JOP (full)                79,685 │        1,007 (1.26%)            276 (0.35%) 
├─────────────┼──────────────────────┼──────────────────────┼───────────────────────┤
  JOP (part)                     - │      16,458 (20.65%)        12,461 (15.64%) 
├─────────────┼──────────────────────┼──────────────────────┼───────────────────────┤
  SYS (full)                 8,276 │          422 (5.10%)            119 (1.44%) 
├─────────────┼──────────────────────┼──────────────────────┼───────────────────────┤
  SYS (part)                     - │       4,317 (52.16%)         3,864 (46.69%) 
└─────────────┴──────────────────────┴──────────────────────┴───────────────────────┘

Note these totals exclude low-quality gadgets (use --all flag to include). In the output table, we see that up to 72.02% of individual ROP gadgets, and 15.64% of JOP gadgets, are portable across all three versions (counting partial matches).

Acknowledgements

This project started as an optimized solution to Chapter 8, exercise 3 of "Practical Binary Analysis" by Dennis Andreisse (affiliate link), and builds on the design outlined therein.

Free book about software assurance: https://highassurance.rs/

License and Contributing

Licensed under the MIT license. Contributions are welcome!

Dependencies

~16–27MB
~449K SLoC