Automated tools produce vulnerability reports at unprecedented scale, but turning a report into a confirmed proof-of-vulnerability remains challenging. Directed grey-box fuzzers are widely used for validation but lack the ability to consume the rich context these reports carry, reducing each target to a bare location. In contrast, LLM-based approaches can consume this context but are often unable to identify the true triggering mechanism or to synthesize a working proof-of-vulnerability. We present CtxFuzz, a context-aware directed fuzzer that enables an LLM agent and a coverage-guided fuzzer to iteratively exchange and refine bug-triggering context throughout the campaign. The fuzzer contributes execution context, such as what inputs were tried and how the program behaved, grounding the agent’s reasoning in concrete evidence. The agent contributes semantic context, including hypotheses about the triggering conditions, explanations of failed attempts, and targeted mutations to overcome them, guiding the fuzzer toward relevant regions of the input space. Failed attempts are not discarded but decomposed into mutation scripts that translate each identified blocker into a fuzzer-executable strategy. This iterative exchange enables CtxFuzz to trigger vulnerabilities that neither directed fuzzers nor LLM agents reach independently. On 62 vulnerabilities across 16 open-source projects from the AIxCC Final competition (39 C, 23 Java), CtxFuzz exposes 61 of 62 targets, including 29 that no state-of-the-art directed fuzzer reaches, and exposes two zero-day vulnerabilities through a real-world vulnerability report validation pipeline.

@inproceedings{paperbot12590,
  title = {CtxFuzz: Context-Aware Directed Fuzzing via Runtime Fuzzer-Agent Cooperation},
  author = {Jiho Kim and Dongkwan Kim and HyungSeok Han and Youngjoon Kim and Soyeon Park and Sangwoo Ji and Joshua Wang and Dae R. Jeong and Taesoo Kim},
  booktitle = {ASE},
  year = {2026},
  doi = {10.1145/3832783.3834365}
}