By Michael D. Brown

We’re excited to share that Trail of Bits has been selected as one of the seven exclusive teams to participate in the small business track for DARPA’s AI Cyber Challenge (AIxCC). Our team will receive a $1 million award to create a Cyber Reasoning System (CRS) and compete in the AIxCC Semifinal Competition later this summer. This recognition not only highlights our dedication to advancing cybersecurity but also marks a significant milestone in our journey in pioneering solutions that could shape the future of AI-driven security. Our involvement in the AIxCC represents a step forward in our commitment to pushing the boundaries of what’s possible, envisioning a future where cybersecurity challenges are met with innovative, AI-powered solutions.

As we move beyond the initial phase of the competition, we’re eager to offer a sneak peek into the driving forces behind our approach, without spilling all of our secrets, of course. In a field where competitors often hold their cards close to their chests, we at Trail of Bits believe in the value of openness and sharing. Our motivation stems from more than just the desire to compete; it’s about contributing to a broader understanding and development within the cybersecurity community. While we navigate through this challenge with an eye on victory, our aim is also to foster a culture of transparency and collaboration, aligning with our deep-rooted open-source ethos.

For background on the challenge, see our two previous posts on the AIxCC:

• DARPA’s AI Cyber Challenge: We’re In!
• Our thoughts on AIxCC’s competition format

*** Disclaimer: Information about AIxCC’s rules, structure, and events referenced in this document are subject to change. This post is NOT an authoritative document. Please refer to DARPA’s website and official documents for first-hand information. ***

The guiding principles for building our CRS

In addition to competing in the AIxCC’s spiritual predecessor, the Cyber Grand Challenge (CGC), our team at Trail of Bits has been working to apply AI/ML techniques to critical cybersecurity problems for many years. These experiences have heavily influenced our approach to the AIxCC. While we’ll be waiting until later in the competition to share specific details, we would like to share the guiding principles for building our AI/ML-driven CRS that have come from this work:

CRS architecture is key to achieving scalability, resiliency, and versatility

DARPA’s CGC, like the AIxCC, tasked competitors with developing CRSs that find vulnerabilities at scale (i.e., that scan many challenge programs in a limited period of time) without any human intervention. The CRS Trail of Bits created to compete in the CGC, Cyberdyne, addressed these problems with a distributed system architecture. Cyberdyne provisioned many independent nodes, each capable of performing key tasks such as fuzzing and symbolic execution. Each node was tasked with one or more challenge problems, and could even cooperate with other nodes on the same challenge.

This design had several advantages. First, the CRS maximized coverage of the 131 challenges via parallel processing. This allowed the CRS to both achieve the scale needed to succeed in the competition and avoid being bogged down with particularly challenging problems. Second, the CRS was resilient to localized failures. If nodes experienced a catastrophic error while analyzing a challenge problem, the operation of other independent nodes was not affected, limiting the damage to the CRS’s overall score. The care taken in this design paid off in the competition: Cyberdyne ranked second among all CRSs in terms of the total number of verified bugs found!

The format of the AIxCC bears a strong resemblance to that of the CGC, so the CRS we build for the AIxCC will also need to be scalable and resilient to failures. However, the AIxCC has an additional wrinkle—challenge diversity. The AIxCC’s challenge problem set will include programs written in languages other than C/C++, including many interpreted languages such as Java and Python. This will require a successful CRS to be highly versatile. Fortunately, the distributed architecture used in Cyberdyne can be adapted for the AIxCC to address versatility in a manner similar to scalability and resiliency. The key difference is that problem-solving nodes used for AIxCC challenges will need to be specialized for different types of challenge problems.

AI/ML is best for complementing conventional techniques, not replacing them

I, along with my co-authors from Georgia Tech, recently presented work at the USENIX Security Symposium on an ML-based static analysis tool we built called VulChecker. VulChecker uses graph-based ML algorithms to locate and classify vulnerabilities in program source code. We evaluated VulChecker against a commercial static analysis tool and found that VulChecker outperformed the commercial tool at detecting certain vulnerability types that rule-based tools typically struggle with, such as integer overflow/underflow vulnerabilities. However, for vulnerabilities that are amenable to rule-based checks (e.g., stack buffer overflow vulnerabilities), VulChecker was effective but did not outperform conventional static analysis.

Considering that rule-based checks are generally less costly to implement than ML models, it doesn’t make sense to replace conventional analysis entirely with AI/ML. Rather, AI/ML is best suited to complement conventional approaches by addressing the problem instances that they struggle with. In the context of the AIxCC, our experience suggests that an AI/ML-only approach is a losing proposition due to high compute costs and the effect of compounding false positives, inaccuracies, and/or confabulations at each step. With that in mind, we plan to use AI/ML in our CRS only where it is best suited or where no conventional options exist. For now, we are planning to use AI/ML approaches primarily for vulnerability detection/classification, patch generation, and input generation tasks in our CRS.

Use the right AI/ML models for the job!

LLMs have been demonstrated to have many emergent capabilities due to the sheer size of their training sets. Among the tasks a CRS must complete in the AIxCC that are suitable for AI/ML, several are tailor-made for LLMs, such as generating code snippets and seed inputs for fuzzing. However, based on our past research, we’ve found that LLMs may not actually be the best option for such tasks.

Last fall, our team supported the United Kingdom’s Frontier AI Taskforce’s efforts to evaluate the risks posed by frontier AI models. We created a framework for rigorously assessing the offensive cyber capabilities of LLMs, which allowed us to 1) rate the model’s independent capabilities relative to human skill levels (i.e., novice, intermediate, expert) and 2) rate the model’s ability to upskill a novice or intermediate human operator. We used this framework to assess different LLMs’ abilities to handle several distinct tasks, including those highly relevant to AIxCC (e.g., vulnerability discovery and contextualization).

We found that LLMs could perform only as well as experts or significantly upskill novices for tasks that were reducible to natural language processing, such as writing phishing emails and conducting misinformation campaigns. For other cyber tasks (including those relevant to the AIxCC) such as creating malicious software, finding vulnerabilities in source code, and creating exploits, current-generation LLMs had novice-like capabilities and could only marginally upskill novice users. These results speak to the lack of reasoning and planning capabilities in LLMs, which has been well documented.

Because LLMs will struggle greatly with tasks that are reasoning-intensive, such as identifying novel instances of vulnerabilities in source code or classifying vulnerabilities, we’ll avoid their use in our CRS. Other types of AI/ML models with narrower scopes are a better option. Expecting LLMs to perform well on these tasks risks high levels of inaccuracy or false positives that can derail late tasks (e.g., generating patches).

What’s next?

Next month, DARPA will hold its AIxCC kickoff event where we should learn more about the infrastructure DARPA will provide for the competition. Once released, we expect this information will allow us (and other competing teams) to make more concrete progress toward building our CRS.

*** This is a Security Bloggers Network syndicated blog from Trail of Bits Blog authored by Trail of Bits. Read the original post at: https://blog.trailofbits.com/2024/03/11/darpa-awards-1-million-to-trail-of-bits-for-ai-cyber-challenge/