Ruijie Meng

Ruijie Meng, Gregory J. Duck, Abhik Roychoudhury

Greybox fuzzing is one of the most popular methods for detecting software vulnerabilities, which conducts a biased random search within the program input space. To enhance its effectiveness in achieving deep coverage of program behaviors, greybox fuzzing is often combined with concolic execution, which performs a path-sensitive search over the domain of program inputs. In hybrid fuzzing, conventional greybox fuzzing is followed by concolic execution in an iterative loop, where reachability roadblocks encountered by greybox fuzzing are tackled by concolic execution. However, such hybrid fuzzing still suffers from difficulties conventionally faced by concolic execution, such as the need for environment modeling and system call support. In this work, we explore the potential of developing "smart" concolic execution empowered by Large Language Models (LLMs), leveraging their knowledge of code semantics during constraint computing and solving. When coverage-based greybox fuzzing reaches a roadblock in terms of reaching certain branches, we conduct a slicing on the execution trace and suggest modifications of the input to reach the relevant branches. The LLM is used as a solver to generate the modified input to reach the desired branches. Compared with state-of-the-art hybrid fuzzers CoFuzz, Intriguer, and QSYM, our LLM-based hybrid fuzzer HyllFuzz(pronounced "hill fuzz") covers 31.43%, 44.56%, and 59.48% more code branches, respectively. Furthermore, the LLM-based concolic execution in HyllFuzz takes a time that is 3--19 times faster than the concolic execution running in existing hybrid fuzzing tools. In extensively tested real-world subjects, HyllFuzz exposed seven previously unknown bugs. This experience shows that LLMs can be effectively inserted into the iterative loop of hybrid fuzzers to efficiently expose more program behaviors.

Haoxin Tu, Huan Zhao, Yahui Song et al.

Automatically generated code is gaining traction recently, owing to the prevalence of Large Language Models (LLMs). Further, the AlphaProof initiative has demonstrated the possibility of using AI for general mathematical reasoning. Reasoning about computer programs (software) can be accomplished via general mathematical reasoning; however, it tends to be more structured and richer in contexts. This forms an attractive proposition, since then AI agents can be used to reason about voluminous code that gets generated by AI. In this work, we present a first LLM agent, AutoRocq, for conducting program verification. Unlike past works, which rely on extensive training of LLMs on proof examples, our agent learns on-the-fly and improves the proof via an iterative refinement loop. The iterative improvement of the proof is achieved by the proof agent communicating with the Rocq (formerly Coq) theorem prover to get additional context and feedback. The final result of the iteration is a proof derivation checked by the Rocq theorem prover. In this way, our proof construction involves autonomous collaboration between the proof agent and the theorem prover. This autonomy facilitates the search for proofs and decision-making in deciding on the structure of the proof tree. Experimental evaluation on SV-COMP benchmarks and on Linux kernel modules shows promising efficacy in achieving automated program verification. As automation in code generation becomes more widespread, we posit that our proof agent can be potentially integrated with AI coding agents to achieve a generate and validate loop, thus moving closer to the vision of trusted automatic programming.

Yuntong Zhang, Sungmin Kang, Ruijie Meng et al.

Agentic AI has been a topic of great interest recently. A Large Language Model (LLM) agent involves one or more LLMs in the back-end. In the front end, it conducts autonomous decision-making by combining the LLM outputs with results obtained by invoking several external tools. The autonomous interactions with the external environment introduce critical security risks. In this paper, we present a grey-box approach to explore diverse behaviors and uncover security risks in LLM agents. Our approach VeriGrey uses the sequence of tools invoked as a feedback function to drive the testing process. This helps uncover infrequent but dangerous tool invocations that cause unexpected agent behavior. As mutation operators in the testing process, we mutate prompts to design pernicious injection prompts. This is carefully accomplished by linking the task of the agent to an injection task, so that the injection task becomes a necessary step of completing the agent functionality. Comparing our approach with a black-box baseline on the well-known AgentDojo benchmark, VeriGrey achieves 33% additional efficacy in finding indirect prompt injection vulnerabilities with a GPT-4.1 back-end. We also conduct real-world case studies with the widely used coding agent Gemini CLI, and the well-known OpenClaw personal assistant. VeriGrey finds prompts inducing several attack scenarios that could not be identified by black-box approaches. In OpenClaw, by constructing a conversation agent which employs mutational fuzz testing as needed, VeriGrey is able to discover malicious skill variants from 10 malicious skills (with 10/10= 100% success rate on the Kimi-K2.5 LLM backend, and 9/10= 90% success rate on Opus 4.6 LLM backend). This demonstrates the value of a dynamic approach like VeriGrey to test agents, and to eventually lead to an agent assurance framework.

Zihao Xu, Xiao Cheng, Ruijie Meng et al.

LLM API calls are becoming a ubiquitous program construct, yet they create a boundary that no existing program analysis can cross: runtime values enter a natural-language prompt, undergo opaque processing inside the LLM, and re-emerge as code, SQL, JSON, or text that the program consumes. Every analysis that tracks data across function boundaries, including taint analysis, program slicing, dependency analysis, and change-impact analysis, relies on dataflow summaries of callee behavior. LLM calls have no such summaries, breaking all of these analyses at what we call the NL/PL boundary. We present the first information flow method to bridge this boundary. Grounded in quantitative information flow theory, our taxonomy defines 24 labels along two orthogonal dimensions: information preservation level (from lexically preserved to fully blocked) and output modality (natural language, structured format, executable artifact). We label 9,083 placeholder-output pairs from 4,154 real-world Python files and validate reliability with Cohen's $κ= 0.82$ and near-complete coverage (0.01\% unclassifiable). We demonstrate the taxonomy's utility on two downstream applications: (1)~a two-stage taint propagation pipeline combining taxonomy-based filtering with LLM verification achieves $F_1 = 0.923$ on 353 expert-annotated pairs, with cross-language validation on six real-world OpenClaw prompt injection cases further confirming effectiveness; (2)~taxonomy-informed backward slicing reduces slice size by a mean of 15\% in files containing non-propagating placeholders. Per-label analysis reveals that four blocked labels account for nearly all non-propagating cases, providing actionable filtering criteria for tool builders.

Kaitong Zheng, Ruijie Meng, Chengshi Zheng et al.

With the development of deep neural networks, automatic music composition has made great progress. Although emotional music can evoke listeners' different emotions and it is important for artistic expression, only few researches have focused on generating emotional music. This paper presents EmotionBox -an music-element-driven emotional music generator that is capable of composing music given a specific emotion, where this model does not require a music dataset labeled with emotions. Instead, pitch histogram and note density are extracted as features that represent mode and tempo respectively to control music emotions. The subjective listening tests show that the Emotionbox has a more competitive and balanced performance in arousing a specified emotion than the emotion-label-based method.

Ruijie Meng, Zhen Dong, Jialin Li et al.

Software model checking is a verification technique which is widely used for checking temporal properties of software systems. Even though it is a property verification technique, its common usage in practice is in "bug finding", that is, finding violations of temporal properties. Motivated by this observation and leveraging the recent progress in fuzzing, we build a greybox fuzzing framework to find violations of Linear-time Temporal Logic (LTL) properties. Our framework takes as input a sequential program written in C/C++, and an LTL property. It finds violations, or counterexample traces, of the LTL property in stateful software systems; however, it does not achieve verification. Our work substantially extends directed greybox fuzzing to witness arbitrarily complex event orderings. We note that existing directed greybox fuzzing approaches are limited to witnessing reaching a location or witnessing simple event orderings like use-after-free. At the same time, compared to model checkers, our approach finds the counterexamples faster, thereby finding more counterexamples within a given time budget. Our LTL-Fuzzer tool, built on top of the AFL fuzzer, is shown to be effective in detecting bugs in well-known protocol implementations, such as OpenSSL and Telnet. We use LTL-Fuzzer to reproduce known vulnerabilities (CVEs), to find 15 zero-day bugs by checking properties extracted from RFCs (for which 12 CVEs have been assigned), and to find violations of both safety as well as liveness properties in real-world protocol implementations. Our work represents a practical advance over software model checkers -- while simultaneously representing a conceptual advance over existing greybox fuzzers. Our work thus provides a starting point for understanding the unexplored synergies between software model checking and greybox fuzzing.