“Estragon: I'm like that. Either I forget right away or I never forget.” - Samuel Beckett, Waiting for Godot

Hacking and Automation

As hackers, we spend a lot of time making things easier for ourselves.

For example, you might be aware of a tool called Metasploit, which can be used to make getting into a target easier. We’ve also built internet-scale scanning tools, allowing us to easily view data about open ports across the internet. Some of our less ethical comrades-in-arms build worms and botnets to automate the process of doing whatever they want to do.

If the heart of hacking is making things do what they shouldn’t, then perhaps the lungs are automation.

Over the years, we’ve seen security in general and vulnerability discovery in particular move from a risky, shady business to massive corporate-sponsored activities with open marketplaces for bug bounties. We’ve also seen a concomitant improvement in the techniques of hacking.

If hackers had known in 1996 that we’d go from Stack-based buffer overflows to chaining ROP gadgets, perhaps we’d have asserted “no free bugs” earlier on. This maturity has allowed us to find a number of bugs that would have been unbelievable in the early 2000s, and exploits for those bugs are quickly packaged into tools like Metasploit.

Now that we’ve automated the process of running our exploits once they’ve been written, why is it so hard to get machines to find the bugs for us?

This is, of course, not for lack of trying. Fuzzing is a powerful technique that turns up a ton of bugs in an automated way. In fact, fuzzing is powerful enough that loads of folks turn up 0-days while they’re learning how to do fuzzing!

However, the trouble with fuzzing is that you never know what you’re going to turn up, and once you get a crash, there is a lot of work left to be done to craft an exploit and understand how and why the crash occurred — and that’s on top of all the work needed to craft a reliable exploit.

Automated bug finding, like we saw in the DARPA Cyber Grand Challenge, takes this to another level by combining fuzzing and symbolic execution with other program analysis techniques, like reachability and input dependence. But fuzzers and SMT solvers — a program that solves particular types of logic problems — haven’t found all the bugs, so what are we missing?

As with many problems in the last few years, organizations are hoping the answer lies in artificial intelligence and machine learning. The trouble with this hope is that AI is good at some tasks, and bug finding may simply not be one of them — at least not yet.

Learning to Find Bugs

Academic literature is rich with papers aiming to find bugs with machine learning. A quick Google Scholar search turns up over 140,000 articles on the topic as of this writing, and many of these articles seem to promise that, any day now, machine learning algorithms will turn up bugs in your source code.

There are a number of developer tools that suggest this could be true. Tools like Codota, Tabnine, and Kite will help auto-complete your code and are quite good. In fact, Microsoft has used GPT-3 to write code from natural language.

But creating code and finding bugs is sadly an entirely different problem. A 2017 paper written by Chappell et al — a collaboration between Australia’s Queensland University of Technology and the technology giant Oracle — found that a variety of machine learning approaches vastly underperformed Oracle’s Parfait system, which uses more traditional symbolic analysis techniques on the intermediate representations used by the compiler.

Another paper, out of the University of Oslo in Norway, Simulated SQL injection using Q-Learning, a form of Reinforcement Learning. This paper caused a stir in the MLSec community and especially within the DEF CON AI village (full disclosure: I am an officer for the AI village and helped cause the stir). The possibility of using a roomba-like method to find bugs was deeply enticing, and Erdodi et al. did great work.

However, their method requires a very particular environment, and although the agent learned to exploit the specific simulation, the method does not seem to generalize well. So, what’s a hacker to do?

Blaming Our Boots

“Vladimir: There’s man all over for you, blaming on his boots the faults of his feet.” - Samuel Beckett, Waiting for Godot

One of the fundamental problems with throwing machine learning at security problems is that many ML techniques have been optimized for particular types of data. This is particularly important for deep learning techniques.

Images are tensors — a sort of matrix with not just height and width but also color channels — of rectangular bit maps with a range of possible values for each pixel. Natural language is tokenized, and those tokens are mapped into a word embedding, like GloVe or Word2Vec.

This is not to downplay the tremendous accomplishments of these machine learning techniques but to demonstrate that, in order for us to repurpose them, we must understand why they were built this way.

Unfortunately, the properties we find important for computer vision — shift invariance and orientation invariance — are not properties that are important for tasks like vulnerability detection or malware analysis. There is, likewise, a heavy dependence in log analysis and similar tasks on tokens that are unlikely to be in our vocabulary — oddly encoded strings, weird file names, and unusual commands. This makes these techniques unsuitable for many of our defensive tasks and, for similar reasons, mostly useless for generating net-new exploits.

Why doesn’t this work? A few issues are at play here. First, the machine does not understand what it is learning. Machine learning algorithms are ultimately function approximators — systems that see some inputs and some outputs, and figure out what function generated them.

For example, if our dataset is:

X = {1, 3, 7, 11, 2}

Y = {3, 7, 15, 23, 5}

Our algorithm might see the first input and output: 3 = f(1) and guess that f(x) = 3x.

By the second input, it would probably be able to figure out that y = f(x) = 2x + 1.

By the fifth, there would be a lot of good evidence that f(x) = 2x + 1. But this is a simple linear model, with one weight term and one bias term. Once we have to account for a large number of dimensions and a function that turns a label like “cat” into a 32 x 32 image with 3 color channels, approximating that function becomes much harder.

It stands to reason then that the function which maps a few dozen lines of code that spread across several files into a particular class of vulnerability will be harder still to approximate.

Ultimately, the problem is neither the technology on its own nor the data representation on its own. It is that we are trying to use the data we have to solve a hard problem without addressing the underlying difficulties of that problem.

In our case, the challenge is not identifying vulnerabilities that look like others we’ve found before. The challenge is in capturing the semantic meaning of the code and the code flow at a point, and using that information to generate an output that tells us whether or not a certain condition is met.

This is what SAT solvers are trying to do. It is worth noting then that, from a purely theoretical perspective, this is the problem SAT, the canonical NP-Complete problem. It explains why the problem is so hard — we’re trying to solve one of the most challenging problems in computer science!

Waiting for Godot

The Samuel Beckett play, Waiting for Godot, centers around the characters of Vladimir and Estragon. The two characters are, as the title suggests, waiting for a character named Godot. To spoil a roughly 70-year-old play, I’ll give away the punchline: Godot never comes.

Today, security researchers who are interested in using artificial intelligence and machine learning to move the ball forward are in a similar position. We sit or stand by the leafless tree, waiting for our AI Godot. Like Vladimir and Estragon, our Godot will never come if we wait.

If we want to see more automation and applications of machine learning to vulnerability discovery, it will not suffice to repurpose convolutional neural networks, gradient-boosted decision trees, and transformers. Instead, we need to think about the way we represent data and how to capture the relevant details of that data. Then, we need to develop algorithms that can capture, learn, and retain that information.

We cannot wait for Godot — we have to find him ourselves.