Answers are grounded in this paper and cite their sources.
Fuzzing is a widely adopted software testing technique in software engineering for discovering bugs and security vulnerabilities. However, even state-of-the-art mutation-based or generation-based fuzzers fail to provide efficient and scalable solutions for real-world applications that expect highly structured file formats. In this paper, we present AdapFuzz, a novel adaptive fuzzing highlight>scheme that attempts to utilize a large language model (LLM) as a generator or mutator to effectively yield high-quality seeds or test inputs. More specifically, we devise various prompts to instruct the LLM to generate new seeds or modify existing ones. Furthermore, we propose a valid-byte-aware seed mutation strategy that leverages coverage-sensitive probe information to guide the deterministic and non-deterministic mutation stages of the fuzzer, making our scheme more efficient. We implement this scheme on top of state-of-the-art AFL++, AFL, FairFuzz, and MOPT. Our evaluations on nine real-world applications show that AdapFuzz increases the number of paths explored and crashes discovered by 14.69% and 107.19%, respectively, over AFL++. Further comparison demonstrates that AdapFuzz outperforms other advanced fuzzers, e.g., FairFuzz and MOPT. Also, extensive evaluations indicate that AdapFuzz has excellent effectiveness, compatibility, and scalability.
No takes yet. Share an insight, caveat, or question.
Wang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: