arXiv subjects

Sicong Cao

Publications and source records attributed to Sicong Cao.

13 recordsLinked to original sources

arXiv2026

DreamGuard: Efficient Runtime Guardrail for LLM Agents via Risk-Aware World Model

As large language model (LLM) agents increasingly invoke external tools and interact with real-world systems, unsafe actions may cause irreversible consequences on external states, user data, and downstream services. Recent runtime guardrails mitigate such risks by checking proposed actions before execution, but many remain reactive: they primarily assess the apparent safety of the current action, lacking an explicit model of how risk evolves across the trajectory. This limitation creates a critical blind spot for long-horizon risks, where individually benign-looking actions can gradually drift the agent toward hazardous states. In response, we propose DreamGuard, a proactive guardrail for LLM agents built around a risk-aware world model. The world model maintains a compact recurrent latent state over the trajectory and predicts future latent states from which DreamGuard derives immediate-hazard and prefix-risk evidence. It then fuses these multi-horizon signals into intervention decisions before execution. Experiments across four benchmarks and an online guardrail evaluation show that DreamGuard outperforms generic, reactive, and proactive guardrail baselines, achieves the best safety-utility trade-off among evaluated guardrails, and maintains an average end-to-end latency of 25 ms per call.

cs.AI↗

arXiv2026

Knowledge-Enhanced Agentic Vulnerability Repair

Frontier foundation models have changed the math on vulnerability discovery, but the bigger challenge is how the remediation side keeps up. Despite recent progresses in Automated Vulnerability Repair (AVR), current solutions struggle to reliably identify the root causes of vulnerabilities, and insufficiently utilize the prior fix knowledge to guide the patch generation process, thus undermining their effectiveness in practice. To address this gap, we propose KeaRepair, a novel agentic AVR approach that grounds patch generation in verified program facts and high-level vulnerability knowledge. Specifically, KeaRepair first extracts multi-dimensional vulnerability knowledge from historical vulnerability-patch pairs from dual complementary views, and constructs dedicated retrieval knowledge bases. It then employs a tool-augmented agent that performs ReAct-style reasoning to collect verified program facts for vulnerability diagnosis. Finally, based on the diagnostic results, KeaRepair performs knowledge-level retrieval-augmented patch generation and iteratively refines patches through a closed-loop validation process involving compilation, PoC replay, and test-suite execution. Experimental results show that KeaRepair significantly outperforms existing AVR approaches on 55 reproducible C/C++ vulnerabilities. When paired with Gemini-3.1-Pro, KeaRepair successfully repairs 46 vulnerabilities, achieving a repair rate of 83.64%. Moreover, KeaRepair fixes six unique vulnerabilities that none of the baselines can address, and further demonstrates strong cross-language generalizability.

cs.SE↗

arXiv2026

MAS-SZZ: Multi-Agentic SZZ Algorithm for Vulnerability-Inducing Commit Identification

Accurate vulnerability-inducing commit identification serves as a foundation for a series of software security tasks, such as vulnerability detection and affected version analysis. A straightforward solution is the SZZ algorithm, which traces back through the code history to identify the earliest commit that modify the vulnerable code. Unfortunately, neither the customized V-SZZ nor state-of-the-art LLM4SZZ perform satisfactorily due to the incorrect anchor selection and inadequate backtracking capability, making them far beyond a reliable usage in practice. To overcome these challenges, we propose a multi-agentic SZZ algorithm, named MAS-SZZ, that facilitates the identification of vulnerability-inducing commits through collaboration among agents. Specifically, given a CVE description and its corresponding fixing commit, MAS-SZZ summarizes the root cause of the vulnerability and employs a structured step-forward prompting strategy to localize vulnerability-related statements based on the change intent of each patch hunk. These vulnerable statements serve as anchors from which MAS-SZZ autonomously traces backward through the repository's history to find the commit that first introduced the vulnerability. Extensive experiments show that MAS-SZZ outperforms the state-of-the-art baselines across datasets and programming languages, achieving F1-score gains of up to 65.22% over the best-performing SZZ algorithm.

cs.CR↗

arXiv2026

ICON: Intent-Context Coupling for Efficient Multi-Turn Jailbreak Attack

Multi-turn jailbreak attacks have emerged as a critical threat to Large Language Models (LLMs), bypassing safety mechanisms by progressively constructing adversarial contexts from scratch and incrementally refining prompts. However, existing methods suffer from the inefficiency of incremental context construction that requires step-by-step LLM interaction, and often stagnate in suboptimal regions due to surface-level optimization. In this paper, we characterize the Intent-Context Coupling phenomenon, revealing that LLM safety constraints are significantly relaxed when a malicious intent is coupled with a semantically congruent context pattern. Driven by this insight, we propose ICON, an automated multi-turn jailbreak framework that efficiently constructs an authoritative-style context via prior-guided semantic routing. Specifically, ICON first routes the malicious intent to a congruent context pattern (e.g., Scientific Research) and instantiates it into an attack prompt sequence. This sequence progressively builds the authoritative-style context and ultimately elicits prohibited content. In addition, ICON incorporates a Hierarchical Optimization Strategy that combines local prompt refinement with global context switching, preventing the attack from stagnating in ineffective contexts. Experimental results across eight SOTA LLMs demonstrate the effectiveness of ICON, achieving a state-of-the-art average Attack Success Rate (ASR) of 97.1\%. Code is available at https://github.com/xwlin-roy/ICON.

cs.CR↗

arXiv2025

Well Begun is Half Done: Location-Aware and Trace-Guided Iterative Automated Vulnerability Repair

The advances of large language models (LLMs) have paved the way for automated software vulnerability repair approaches, which iteratively refine the patch until it becomes plausible. Nevertheless, existing LLM-based vulnerability repair approaches face notable limitations: 1) they ignore the concern of locations that need to be patched and focus solely on the repair content. 2) they lack quality assessment for generated candidate patches in the iterative process. To tackle the two limitations, we propose \sysname, an LLM-based approach that provides information about where should be patched first. Furthermore, \sysname improves the iterative repair strategy by assessing the quality of test-failing patches and selecting the best patch for the next iteration. We introduce two dimensions to assess the quality of patches: whether they introduce new vulnerabilities and the taint statement coverage. We evaluated \sysname on a real-world C/C++ vulnerability repair dataset VulnLoc+, which contains 40 vulnerabilities and their Proofs-of-Vulnerability. The experimental results demonstrate that \sysname exhibits substantial improvements compared with the Neural Machine Translation-based, Program Analysis-based, and LLM-based state-of-the-art vulnerability repair approaches. Specifically, \sysname is able to generate 27 plausible patches, which is comparable to or even 8 to 22 more plausible patches than the baselines. In terms of correct patch generation, \sysname repairs 8 to 13 additional vulnerabilities compared with existing approaches.

cs.SE↗

arXiv2025

Federated Unlearning in the Wild: Rethinking Fairness and Data Discrepancy

Machine unlearning is critical for enforcing data deletion rights like the "right to be forgotten." As a decentralized paradigm, Federated Learning (FL) also requires unlearning, but realistic implementations face two major challenges. First, fairness in Federated Unlearning (FU) is often overlooked. Exact unlearning methods typically force all clients into costly retraining, even those uninvolved. Approximate approaches, using gradient ascent or distillation, make coarse interventions that can unfairly degrade performance for clients with only retained data. Second, most FU evaluations rely on synthetic data assumptions (IID/non-IID) that ignore real-world heterogeneity. These unrealistic benchmarks obscure the true impact of unlearning and limit the applicability of current methods. We first conduct a comprehensive benchmark of existing FU methods under realistic data heterogeneity and fairness conditions. We then propose a novel, fairness-aware FU approach, Federated Cross-Client-Constrains Unlearning (FedCCCU), to explicitly address both challenges. FedCCCU offers a practical and scalable solution for real-world FU. Experimental results show that existing methods perform poorly in realistic settings, while our approach consistently outperforms them.

cs.LG↗

arXiv2025

MalGuard: Towards Real-Time, Accurate, and Actionable Detection of Malicious Packages in PyPI Ecosystem

Malicious package detection has become a critical task in ensuring the security and stability of the PyPI. Existing detection approaches have focused on advancing model selection, evolving from traditional machine learning (ML) models to large language models (LLMs). However, as the complexity of the model increases, the time consumption also increases, which raises the question of whether a lightweight model achieves effective detection. Through empirical research, we demonstrate that collecting a sufficiently comprehensive feature set enables even traditional ML models to achieve outstanding performance. However, with the continuous emergence of new malicious packages, considerable human and material resources are required for feature analysis. Also, traditional ML model-based approaches lack of explainability to malicious packages.Therefore, we propose a novel approach MalGuard based on graph centrality analysis and the LIME (Local Interpretable Model-agnostic Explanations) algorithm to detect malicious packages.To overcome the above two challenges, we leverage graph centrality analysis to extract sensitive APIs automatically to replace manual analysis. To understand the sensitive APIs, we further refine the feature set using LLM and integrate the LIME algorithm with ML models to provide explanations for malicious packages. We evaluated MalGuard against six SOTA baselines with the same settings. Experimental results show that our proposed MalGuard, improves precision by 0.5%-33.2% and recall by 1.8%-22.1%. With MalGuard, we successfully identified 113 previously unknown malicious packages from a pool of 64,348 newly-uploaded packages over a five-week period, and 109 out of them have been removed by the PyPI official.

cs.CR↗

arXiv2025

A Systematic Literature Review on Explainability for Machine/Deep Learning-based Software Engineering Research

The remarkable achievements of Artificial Intelligence (AI) algorithms, particularly in Machine Learning (ML) and Deep Learning (DL), have fueled their extensive deployment across multiple sectors, including Software Engineering (SE). However, due to their black-box nature, these promising AI-driven SE models are still far from being deployed in practice. This lack of explainability poses unwanted risks for their applications in critical tasks, such as vulnerability detection, where decision-making transparency is of paramount importance. This paper endeavors to elucidate this interdisciplinary domain by presenting a systematic literature review of approaches that aim to improve the explainability of AI models within the context of SE. The review canvasses work appearing in the most prominent SE & AI conferences and journals, and spans 108 papers across 23 unique SE tasks. Based on three key Research Questions (RQs), we aim to (1) summarize the SE tasks where XAI techniques have shown success to date; (2) classify and analyze different XAI techniques; and (3) investigate existing evaluation approaches. Based on our findings, we identified a set of challenges remaining to be addressed in existing studies, together with a set of guidelines highlighting potential opportunities we deemed appropriate and important for future work.

cs.SE↗

arXiv2024

Large Language Model for Vulnerability Detection and Repair: Literature Review and the Road Ahead

The significant advancements in Large Language Models (LLMs) have resulted in their widespread adoption across various tasks within Software Engineering (SE), including vulnerability detection and repair. Numerous studies have investigated the application of LLMs to enhance vulnerability detection and repair tasks. Despite the increasing research interest, there is currently no existing survey that focuses on the utilization of LLMs for vulnerability detection and repair. In this paper, we aim to bridge this gap by offering a systematic literature review of approaches aimed at improving vulnerability detection and repair through the utilization of LLMs. The review encompasses research work from leading SE, AI, and Security conferences and journals, encompassing 43 papers published across 25 distinct venues, along with 15 high-quality preprint papers, bringing the total to 58 papers. By answering three key research questions, we aim to (1) summarize the LLMs employed in the relevant literature, (2) categorize various LLM adaptation techniques in vulnerability detection, and (3) classify various LLM adaptation techniques in vulnerability repair. Based on our findings, we have identified a series of limitations of existing studies. Additionally, we have outlined a roadmap highlighting potential opportunities that we believe are pertinent and crucial for future research endeavors.

cs.SE↗

arXiv2024

Coca: Improving and Explaining Graph Neural Network-Based Vulnerability Detection Systems

Recently, Graph Neural Network (GNN)-based vulnerability detection systems have achieved remarkable success. However, the lack of explainability poses a critical challenge to deploy black-box models in security-related domains. For this reason, several approaches have been proposed to explain the decision logic of the detection model by providing a set of crucial statements positively contributing to its predictions. Unfortunately, due to the weakly-robust detection models and suboptimal explanation strategy, they have the danger of revealing spurious correlations and redundancy issue. In this paper, we propose Coca, a general framework aiming to 1) enhance the robustness of existing GNN-based vulnerability detection models to avoid spurious explanations; and 2) provide both concise and effective explanations to reason about the detected vulnerabilities. \sysname consists of two core parts referred to as Trainer and Explainer. The former aims to train a detection model which is robust to random perturbation based on combinatorial contrastive learning, while the latter builds an explainer to derive crucial code statements that are most decisive to the detected vulnerability via dual-view causal inference as explanations. We apply Coca over three typical GNN-based vulnerability detectors. Experimental results show that Coca can effectively mitigate the spurious correlation issue, and provide more useful high-quality explanations.

cs.CR↗

arXiv2023

ODDFUZZ: Discovering Java Deserialization Vulnerabilities via Structure-Aware Directed Greybox Fuzzing

Java deserialization vulnerability is a severe threat in practice. Researchers have proposed static analysis solutions to locate candidate vulnerabilities and fuzzing solutions to generate proof-of-concept (PoC) serialized objects to trigger them. However, existing solutions have limited effectiveness and efficiency. In this paper, we propose a novel hybrid solution ODDFUZZ to efficiently discover Java deserialization vulnerabilities. First, ODDFUZZ performs lightweight static taint analysis to identify candidate gadget chains that may cause deserialization vulner-abilities. In this step, ODDFUZZ tries to locate all candidates and avoid false negatives. Then, ODDFUZZ performs directed greybox fuzzing (DGF) to explore those candidates and generate PoC testcases to mitigate false positives. Specifically, ODDFUZZ applies a structure-aware seed generation method to guarantee the validity of the testcases, and adopts a novel hybrid feedback and a step-forward strategy to guide the directed fuzzing. We implemented a prototype of ODDFUZZ and evaluated it on the popular Java deserialization repository ysoserial. Results show that, ODDFUZZ could discover 16 out of 34 known gadget chains, while two state-of-the-art baselines only identify three of them. In addition, we evaluated ODDFUZZ on real-world applications including Oracle WebLogic Server, Apache Dubbo, Sonatype Nexus, and protostuff, and found six previously unreported exploitable gadget chains with five CVEs assigned.

cs.CR↗

arXiv2023

Improving Java Deserialization Gadget Chain Mining via Overriding-Guided Object Generation

Java (de)serialization is prone to causing security-critical vulnerabilities that attackers can invoke existing methods (gadgets) on the application's classpath to construct a gadget chain to perform malicious behaviors. Several techniques have been proposed to statically identify suspicious gadget chains and dynamically generate injection objects for fuzzing. However, due to their incomplete support for dynamic program features (e.g., Java runtime polymorphism) and ineffective injection object generation for fuzzing, the existing techniques are still far from satisfactory. In this paper, we first performed an empirical study to investigate the characteristics of Java deserialization vulnerabilities based on our manually collected 86 publicly known gadget chains. The empirical results show that 1) Java deserialization gadgets are usually exploited by abusing runtime polymorphism, which enables attackers to reuse serializable overridden methods; and 2) attackers usually invoke exploitable overridden methods (gadgets) via dynamic binding to generate injection objects for gadget chain construction. Based on our empirical findings, we propose a novel gadget chain mining approach, \emph{GCMiner}, which captures both explicit and implicit method calls to identify more gadget chains, and adopts an overriding-guided object generation approach to generate valid injection objects for fuzzing. The evaluation results show that \emph{GCMiner} significantly outperforms the state-of-the-art techniques, and discovers 56 unique gadget chains that cannot be identified by the baseline approaches.

cs.CR↗

arXiv2022

MVD: Memory-Related Vulnerability Detection Based on Flow-Sensitive Graph Neural Networks

Memory-related vulnerabilities constitute severe threats to the security of modern software. Despite the success of deep learning-based approaches to generic vulnerability detection, they are still limited by the underutilization of flow information when applied for detecting memory-related vulnerabilities, leading to high false positives. In this paper,we propose MVD, a statement-level Memory-related Vulnerability Detection approach based on flow-sensitive graph neural networks (FS-GNN). FS-GNN is employed to jointly embed both unstructured information (i.e., source code) and structured information (i.e., control- and data-flow) to capture implicit memory-related vulnerability patterns. We evaluate MVD on the dataset which contains 4,353 real-world memory-related vulnerabilities, and compare our approach with three state-of-the-art deep learning-based approaches as well as five popular static analysisbased memory detectors. The experiment results show that MVD achieves better detection accuracy, outperforming both state-of-theart DL-based and static analysis-based approaches. Furthermore, MVD makes a great trade-off between accuracy and efficiency.

cs.CR↗