Large Language Model–Based Cyber Reasoning Systems

Abstract

Large language models (LLMs) have made the generation of text and code inexpensive, prompting the question of whether this capability extends to finding and fixing software vulnerabilities. Cyber Reasoning Systems (CRSs), automated systems that detect vulnerabilities, generate inputs that reproduce them, and propose patches, pose this question in its most ambitious form, and the DARPA AI Cyber Challenge (AIxCC) turned it into a measurable competition. <br/>

This dissertation argues that the bottleneck of LLM-based vulnerability discovery is not generating candidate findings but verifying them: an LLM produces plausible reports at negligible cost, and each unverified report shifts work onto the engineer who must triage it. Within the emerging research area of LLM-based CRSs, it presents VulnRep, a system designed, deployed, and evaluated so that no finding reaches a human without an executable input that demonstrably triggers it. VulnRep's architecture realizes this verification-first principle through five decisions: symbol-level rather than file-level code access, Python-based proof-of-vulnerability (PoV) generation, a structural split between a hypothesis-generating Analyzer and a confirming Verifier, sanitizer-based confirmation as the sole verdict mechanism, and the routing of failed PoV attempts into a fuzzing corpus as structurally informed seeds. <br/>

The evaluation combines competition-mode results from the AIxCC environment with controlled component studies and a generalization study across eight open-weight and commercial LLMs. On the AIxCC nginx benchmark, the strongest open-weight model matches the leading commercial model's coverage while trailing it in verification reliability. A comparison against the general-purpose coding agent OpenHands shows that the weaker the underlying model, the larger the margin VulnRep's domain-specific scaffolding provides. This is the strongest evidence in this work that the architecture, rather than any particular model, accounts for the system's performance. Deployed against real-world targets, VulnRep identified ten vulnerabilities across ten distinct OSS-Fuzz projects, two of them previously unreported, with four different LLMs contributing findings. <br/>

Two results generalize beyond VulnRep. First, the strength of a CRS stage is bounded by the strength of the oracle available to it: discovery and reproduction, backed by runtime sanitizers, perform well across models and targets, whereas patch synthesis, lacking a comparably sharp oracle, degrades correspondingly. Second, domain-specific scaffolding makes smaller, locally hosted models viable for CRS workloads, which matters most for deployments restricted by cost, privacy, or air-gap constraints. The verification-first architecture is offered not as a final design but as a baseline for future systems that earn trust through reproducible evidence rather than fluency.