____ __ __ _ _
| _ \ __ _ _ __ ___ _ __ | \/ | __ _ ___| |__ ___| |_ ___ ________
| |_) / _` | '_ \/ _ \ '__| | |\/| |/ _` |/ __| '_ \ / _ \ __/ _ \ /_______/
| __/ (_| | |_)| __/ | | | | | (_| | (__| | | | __/ || __/ \_______\
|_| \__,_| .__/\___|_| |_| |_|\__,_|\___|_| |_|\___|\__\___| /_______/
|_| @==|;;;;;;>
Paper Machete (PM) orchestrates Binary Ninja and Grakn.ai to aid static binary analysis for the purpose of finding bugs in software. PM leverages the Binary Ninja MLIL SSA to extract semantic meaning about individual instructions, operations, register/variable state, and overall control flow.
PM migrates this data into Grakn - a knowledge graph that gives us the ability to define domain-specific ontologies for data and write powerful inference rules to form relationships between data we don't want to (or can't) explicitly store. Heeh, how neat is that?
This project was released in conjunction with a DerbyCon 2017 talk titled "Aiding Static Analysis: Discovering Vulnerabilities in Binary Targets through Knowledge Graph Inferences." You can watch that talk here.
Paper Machete's initial prototype and public codebase were developed by security researchers at the Battelle Memorial Institute. As this project matures, we hope that you will find it useful in your own research and consider contributing to the project.
The BNIL suite of ILs is easy to work with, pleasantly verbose, and human-readable. At any point we can decide to leverage other levels and forms of the IL with little development effort on our part. When you add to that the ability to lift multiple architectures and write custom lifters, we have little reason not to use BNIL.
Grakn's query language (Graql) is easy to learn and intuitive, which is extremely important in the early stages of this research while we're still hand-writing queries to model the patterns vulnerability researchers look for when performing static analysis.
The ability to write our own domain-specific ontologies lets us quickly experiment with new query ideas and ways of making our queries less complex. When we run into a case where we think "gee, if I just had access to the relationship between..." we can modify our ontology and inference rules to get that information.
While the end game for PM is to eliminate the need for human-written queries, the fact is we're starting from square one. Which means hand-jamming a lot queries to model the patterns human vulnerability researchers look for when bug hunting.
Paper Machete requires BinaryNinja v1.1, Grakn v1.4.2, the Grakn Python Driver, and the Java JRE
We've included some basic queries to get you started if you want to play around with PM. As you can imagine, there is no "silver bullet" query that will find all manifestations of a specific vulnerability class. Because of this, we've included versions for each CWE query. As we add new methods of finding the same CWE, we'll add scripts with incremented the version numbers to differentiate.
cwe_120_v1.py - Tests for use of unsafe 'gets()' function (CWE-120)
cwe_121_v1.py - Tests for buffer overflows (CWE-121)
cwe_129_v1.py - Tests for missing bounds checks (CWE-129)
cwe_134_v1.py - Tests for format string vulnerabilities (CWE-134)
cwe_788_v1.py - Tests for missing bounds check on array indexes (CWE-788)
For basic use, run the paper_machete.py script and follow the prompts. For more advanced use, please read the wiki.
Typically you'll start with option [1] and work your way down to option [3]. If you run into any issues with Grakn use option [4] to reset Grakn to a clean state and try again.
... banner ...
[1] Analyze a binary file
[2] Migrate a JSON file into Grakn
[3] Run all CWE queries
[4] Clean and restart Grakn
[5] Quit
Option [1] lists all executable files in the /analysis directory. So place any executables you want to analyze in /analysis. This option will run pmanalyze.py and generate a JSON file in the /analysis directory.
Once you've analyzed files with [1] and produced resulting JSON files, they will appear as a choice in option [2]. Selecting a JSON file in option [2] will migrate the data into Grakn.
Now that you have data in Grakn, you can use option [3]. This will kick off all scripts in /queries against the keyspace of your choice. If you write your own query patterns, just throw them in /queries and option [3] will run them too.