arXiv subjects

Michael Fu

Publications and source records attributed to Michael Fu.

At least 19 recordsLinked to original sources

arXiv2026

XAgent: eXecution-guided Agentic AI for Effective Localization and Resolution of GitHub Issues

Agentic AI has enabled capabilities in leveraging Large Language Models (LLMs) to autonomously resolve repository-level GitHub issues. However, due to the reliance on limited static description of issues, existing agentic approaches suffer from incorrect localization and incomplete validation. Solely relying on this information can bias LLM reasoning toward the narrow scope of the issue description, leading to incomplete patches that fail to address the underlying issue. In this paper, we present XAgent, an execution-guided agentic framework that analyzes dynamic behavior and additional program context to localize and validate issues. The experimental results on the SWE-bench-lite dataset demonstrate that XAgent outperforms other existing approaches, achieving a resolve rate of 62.0% and a function localization accuracy of 72.8%, while maintaining cost efficiency. Our analysis further shows that XAgent successfully resolves 7 additional issues that the top existing baselines fail to address. This work highlights a shift from static, description-oriented patch generation toward dynamic execution-guided issue resolution, opening new opportunities for LLM-based coding agents to achieve more robust and generalizable software maintenance.

cs.SE↗

arXiv2026

Towards a Risk Assessment of Malicious Skill Files in Coding Agents

Autonomous coding agents are increasingly embedded in enterprise software workflows with delegated authority over connected systems. Central to this architecture is the agent skills interface: folders of instructions and scripts that agents load dynamically to specialize their behavior. This interface also widens the attack surface, letting malicious shell commands hide within natural-language skill files. We make three contributions. First, an adversarial skill-synthesis method using six LLMs across four families to transform 471 real-world shell commands into benign-appearing skills, released as a benchmark of 2,826 skills mapped to 11 MITRE ATT&CK tactics. Second, a reproducible evaluation pipeline coupling run stratification, evidence anchoring, a refusal veto, and a deterministic declared-intent override with a three-judge LLM-as-a-judge panel, validated against a blind human gold standard (Cohen's kappa = 0.85). Third, a large-scale characterization of two enterprise-grade agents across 5,629 completed runs. Gemini CLI is exploited in 95.5-96.1% of runs and Qwen Code in 71.6-74.0% (raw majority vote to declared-intent-corrected estimate, both within the human gold standard), nearly invariant to the generating model. Explicit safety recognition occurs in only 1.99% of runs. Enterprises must assess and mitigate skill-interface risk before adopting coding agents. Our code and dataset are available at https://github.com/awsm-research/AgentJailbreak

cs.SE↗

arXiv2026

AgenticRepair: Multi-Faceted Program Context Engineering for Agentic Vulnerability Repair

Automated vulnerability repair aims to reduce the time and effort required to patch security flaws from a vulnerability triage report. Recent agentic AI approaches have shown promising results in automated program repair. However, vulnerability repair demands richer program context than general bug repair - context that security engineers routinely assemble in practice but that existing agentic approaches do not engineer. We identify three critical gaps: code-structure context capturing cross-file data flows and memory operation patterns, runtime-execution context revealing crash semantics and memory origins, and commit-history context recovering how fragile code patterns were introduced. We present AgenticRepair, an agentic vulnerability repair framework that addresses the gaps through multi-faceted program context engineering. AgenticRepair orchestrates three specialized LLM subagents to engineer the contexts, which are then embedded into the memory of a dedicated repair subagent for context-conditioned patch synthesis. Evaluated on SEC-Bench comprising 300 real-world instances with sanitizer-based patch verification, AgenticRepair achieves a 73% success rate, substantially outperforming the strongest baseline by 29%. Our ablation study confirms that the three context facets are mutually complementary, and that multi-agent scaffolding and base-model capacity each play an essential role. Collectively, these findings establish multi-faceted program context engineering as a promising design direction for agentic vulnerability repair.

cs.SE↗

arXiv2026

SkillGate: Cost Efficient Runtime Malicious Skill File Detection in Coding Agents

Software engineering teams now deploy AI coding agents (Cursor, Claude Code, GitHub Copilot) as first-class productivity tools, installing domain-specific skill files to tailor agent behavior to project APIs, framework conventions, and organizational workflows. These complex Markdown files are easily downloaded from public registries with a single npx skills add command and no real security screening, representing a novel supply-chain attack surface: a malicious skill file can silently reprogram agent behavior, exfiltrating credentials, injecting backdoors into generated code, or redirecting agent actions to attacker-controlled endpoints. The threat is not hypothetical: recent reports document hundreds of malicious skill packages in public registries, including organized campaigns that distributed credential-stealing infostealers via fake productivity skills. No systematic toolchain defense exists for this attack surface. We present SkillGate, a deployable security gateway that screens AI skill packages before coding agent installation. SkillGate uses a hybrid regex-prefilter + LLM-judge pipeline: safe-signal files bypass the LLM entirely (skip savings); flagged files have only their matched snippet windows sent to the judge, not the full content (snippet savings). We answer four research questions covering detection effectiveness, screening cost, runtime overhead, and false positive behavior on the SkillsBench benchmark against two existing tools. On SkillsBench (n=1,650, 9.1% malicious), SkillGate achieves F1=0.817, FPR=1.13% while reducing LLM input tokens by 77% vs. full-file screening, and outperforming existing tools by 5-6x on threshold-independent AUPRC (0.830 vs. 0.144/0.162).

cs.SE↗

arXiv2026

AI-Assisted Code Review as a Scaffold for Code Quality and Self-Regulated Learning: An Experience Report

Code review is central to software engineering education but hard to scale in capstone projects due to tight deadlines, uneven peer feedback, and limited prior experience. We investigate an LLM-as-reviewer integrated directly into GitHub pull requests (human-in-the-loop) across two cohorts (more than 100 students, 2023--2024). Using a mixed-methods design -- GitHub data, reflective reports, and a targeted survey -- we examine engagement and responsiveness as behavioral indicators of self-regulated learning processes. Quantitatively, the 2024 cohort produced more iterative activity (1176 vs. 581 PRs), while technical issues observed in 2023 (227 failed AI attempts) dropped to zero after tool and instructional refinements. Despite different adoption levels (93\% vs. 50\% of teams using the tool), responsiveness was stable: 32\% (2023) and 33\% (2024) of successfully AI-reviewed PRs were followed by subsequent commits on the same PR. Qualitatively, students used the LLM's structured comments to focus reviews and discuss code quality, while guidance reduced over-reliance. We contribute: (i) an in-workflow design for an AI reviewer that scaffolds learning while mitigating cognitive offloading; (ii) a repeated cross sectional comparison across two cohorts in authentic settings; (iii) a mixed-methods analysis combining objective GitHub metrics with student self-reports; and (iv) evidence-based pedagogical recommendations for responsible, student-led AI-assisted review.

cs.SE↗

arXiv2026

IntelliSA: An Intelligent Static Analyzer for IaC Security Smell Detection Using Symbolic Rules and Neural Inference

Infrastructure as Code (IaC) enables automated provisioning of large-scale cloud and on-premise environments, reducing the need for repetitive manual setup. However, this automation is a double-edged sword: a single misconfiguration in IaC scripts can propagate widely, leading to severe system downtime and security risks. Prior studies have shown that IaC scripts often contain security smells--bad coding patterns that may introduce vulnerabilities--and have proposed static analyzers based on symbolic rules to detect them. Yet, our preliminary analysis reveals that rule-based detection alone tends to over-approximate, producing excessive false positives and increasing the burden of manual inspection. In this paper, we present IntelliSA, an intelligent static analyzer for IaC security smell detection that integrates symbolic rules with neural inference. IntelliSA applies symbolic rules to over-approximate potential smells for broad coverage, then employs neural inference to filter false positives. While an LLM can effectively perform this filtering, reliance on LLM APIs introduces high cost and latency, raises data governance concerns, and limits reproducibility and offline deployment. To address the challenges, we adopt a knowledge distillation approach: an LLM teacher generates pseudo-labels to train a compact student model--over 500x smaller--that learns from the teacher's knowledge and efficiently classifies false positives. We evaluate IntelliSA against two static analyzers and three LLM baselines (Claude-4, Grok-4, and GPT-5) using a human-labeled dataset including 241 security smells across 11,814 lines of real-world IaC code. Experimental results show that IntelliSA achieves the highest F1 score (83%), outperforming baselines by 7-42%. Moreover, IntelliSA demonstrates the best cost-effectiveness, detecting 60% of security smells while inspecting less than 2% of the codebase.

cs.CR↗

arXiv2025

ASCIIBench: Evaluating Language-Model-Based Understanding of Visually-Oriented Text

Large language models (LLMs) have demonstrated several emergent behaviors with scale, including reasoning and fluency in long-form text generation. However, they continue to struggle with tasks requiring precise spatial and positional reasoning. ASCII art, a symbolic medium where characters encode structure and form, provides a unique probe of this limitation. We introduce ASCIIBench, a novel benchmark for evaluating both the generation and classification of ASCII-text images. ASCIIBench consists of a filtered dataset of 5,315 class-labeled ASCII images and is, to our knowledge, the first publicly available benchmark of its kind. Alongside the dataset, we release weights for a fine-tuned CLIP model adapted to capture ASCII structure, enabling the evaluation of LLM-generated ASCII art. Our analysis shows that cosine similarity over CLIP embeddings fails to separate most ASCII categories, yielding chance-level performance even for low-variance classes. In contrast, classes with high internal mean similarity exhibit clear discriminability, revealing that the bottleneck lies in representation rather than generational variance. These findings position ASCII art as a stress test for multimodal representations and motivate the development of new embedding methods or evaluation metrics tailored to symbolic visual modalities. All resources are available at https://github.com/ASCIIBench/ASCIIBench.

cs.LG↗

arXiv2025

AdaptiveGuard: Towards Adaptive Runtime Safety for LLM-Powered Software

Guardrails are critical for the safe deployment of Large Language Models (LLMs)-powered software. Unlike traditional rule-based systems with limited, predefined input-output spaces that inherently constrain unsafe behavior, LLMs enable open-ended, intelligent interactions--opening the door to jailbreak attacks through user inputs. Guardrails serve as a protective layer, filtering unsafe prompts before they reach the LLM. However, prior research shows that jailbreak attacks can still succeed over 70% of the time, even against advanced models like GPT-4o. While guardrails such as LlamaGuard report up to 95% accuracy, our preliminary analysis shows their performance can drop sharply--to as low as 12%--when confronted with unseen attacks. This highlights a growing software engineering challenge: how to build a post-deployment guardrail that adapts dynamically to emerging threats? To address this, we propose AdaptiveGuard, an adaptive guardrail that detects novel jailbreak attacks as out-of-distribution (OOD) inputs and learns to defend against them through a continual learning framework. Through empirical evaluation, AdaptiveGuard achieves 96% OOD detection accuracy, adapts to new attacks in just two update steps, and retains over 85% F1-score on in-distribution data post-adaptation, outperforming other baselines. These results demonstrate that AdaptiveGuard is a guardrail capable of evolving in response to emerging jailbreak strategies post deployment. We release our AdaptiveGuard and studied datasets at https://github.com/awsm-research/AdaptiveGuard to support further research.

cs.CR↗

arXiv2025

DecipherGuard: Understanding and Deciphering Jailbreak Prompts for a Safer Deployment of Intelligent Software Systems

Intelligent software systems powered by Large Language Models (LLMs) are increasingly deployed in critical sectors, raising concerns about their safety during runtime. Through an industry-academic collaboration when deploying an LLM-powered virtual customer assistant, a critical software engineering challenge emerged: how to enhance a safer deployment of LLM-powered software systems at runtime? While LlamaGuard, the current state-of-the-art runtime guardrail, offers protection against unsafe inputs, our study reveals a Defense Success Rate (DSR) drop of 24% under obfuscation- and template-based jailbreak attacks. In this paper, we propose DecipherGuard, a novel framework that integrates a deciphering layer to counter obfuscation-based prompts and a low-rank adaptation mechanism to enhance guardrail effectiveness against template-based attacks. Empirical evaluation on over 22,000 prompts demonstrates that DecipherGuard improves DSR by 36% to 65% and Overall Guardrail Performance (OGP) by 20% to 50% compared to LlamaGuard and two other runtime guardrails. These results highlight the effectiveness of DecipherGuard in defending LLM-powered software systems against jailbreak attacks during runtime.

cs.SE↗

arXiv2025

On the Evaluation of Large Language Models in Multilingual Vulnerability Repair

Various Deep Learning-based approaches with pre-trained language models have been proposed for automatically repairing software vulnerabilities. However, these approaches are limited to a specific programming language (C/C++). Recent advances in large language models (LLMs) offer language-agnostic capabilities and strong semantic understanding, exhibiting potential to overcome multilingual vulnerability limitations. Although some work has begun to explore LLMs' repair performance, their effectiveness is unsatisfactory. To address these limitations, we conducted a large-scale empirical study to investigate the performance of automated vulnerability repair approaches and state-of-the-art LLMs across seven programming languages. Results show GPT-4o, instruction-tuned with few-shot prompting, performs competitively against the leading approach, VulMaster. Additionally, the LLM-based approach shows superior performance in repairing unique vulnerabilities and is more likely to repair the most dangerous vulnerabilities. Instruction-tuned GPT-4o demonstrates strong generalization on vulnerabilities in previously unseen language, outperforming existing approaches. Analysis shows Go consistently achieves the highest effectiveness across all model types, while C/C++ performs the worst. Based on findings, we discuss the promise of LLM on multilingual vulnerability repair and the reasons behind LLM's failed cases. This work takes the first look at repair approaches and LLMs across multiple languages, highlighting the promising future of adopting LLMs for multilingual vulnerability repair.

cs.SE↗

arXiv2025

SEALGuard: Safeguarding the Multilingual Conversations in Southeast Asian Languages for LLM Software Systems

Safety alignment is critical for LLM-powered systems. While recent LLM-powered guardrail approaches such as LlamaGuard achieve high detection accuracy of unsafe inputs written in English (e.g., ``How to create a bomb?''), they struggle with multilingual unsafe inputs. This limitation leaves LLM systems vulnerable to unsafe and jailbreak prompts written in low-resource languages such as those in Southeast Asia. This paper introduces SEALGuard, a multilingual guardrail designed to improve the safety alignment across diverse languages. It aims to address the multilingual safety alignment gap of existing guardrails and ensure effective filtering of unsafe and jailbreak prompts in LLM-powered systems. We adapt a general-purpose multilingual language model into a multilingual guardrail using low-rank adaptation (LoRA). We construct SEALSBench, a large-scale multilingual safety alignment dataset containing over 260,000 prompts in ten languages, including safe, unsafe, and jailbreak cases. We evaluate SEALGuard against state-of-the-art guardrails such as LlamaGuard on this benchmark. Our findings show that multilingual unsafe and jailbreak prompts substantially degrade the performance of the state-of-the-art LlamaGuard, which experiences a drop in Defense Success Rate (DSR) by 9% and 18%, respectively, compared to its performance on English-only prompts. In contrast, SEALGuard outperforms existing guardrails in detecting multilingual unsafe and jailbreak prompts, improving DSR by 48% over LlamaGuard and achieving the best DSR, precision, and F1-score. Our ablation study further reveals the contributions of adaptation strategies and model size to the overall performance of SEALGuard. We release our pre-trained model and benchmark at https://github.com/awsm-research/SEALGuard to support further research.

cs.CL↗

arXiv2025

Evaluating Large Language Models for Multilingual Vulnerability Detection at Dual Granularities

Various deep learning-based approaches utilizing pre-trained language models (PLMs) have been proposed for automated vulnerability detection. With recent advancements in large language models (LLMs), several studies have begun exploring their application to vulnerability detection tasks. However, existing studies primarily focus on specific programming languages (e.g., C/C++) and function-level detection, leaving the strengths and weaknesses of PLMs and LLMs in multilingual and multi-granularity scenarios largely unexplored. To bridge this gap, we conduct a comprehensive fine-grained empirical study evaluating the effectiveness of state-of-the-art PLMs and LLMs for multilingual vulnerability detection. Using over 30,000 real-world vulnerability-fixing patches across seven programming languages, we systematically assess model performance at both the function-level and line-level. Our key findings indicate that GPT-4o, enhanced through instruction tuning and few-shot prompting, significantly outperforms all other evaluated models, including CodeT5P. Furthermore, the LLM-based approach demonstrates superior capability in detecting unique multilingual vulnerabilities, particularly excelling in identifying the most dangerous and high-severity vulnerabilities. These results underscore the promising potential of adopting LLMs for multilingual vulnerability detection at function-level and line-level, revealing their complementary strengths and substantial improvements over PLM approaches. This empirical evaluation of PLMs and LLMs for multilingual vulnerability detection highlights LLMs' value in addressing real-world software security challenges.

cs.SE↗

arXiv2025

A Preliminary Study of Large Language Models for Multilingual Vulnerability Detection

Deep learning-based approaches, particularly those leveraging pre-trained language models (PLMs), have shown promise in automated software vulnerability detection. However, existing methods are predominantly limited to specific programming languages, restricting their applicability in multilingual settings. Recent advancements in large language models (LLMs) offer language-agnostic capabilities and enhanced semantic understanding, presenting a potential solution to this limitation. While existing studies have explored LLMs for vulnerability detection, their detection performance remains unknown for multilingual vulnerabilities. To address this gap, we conducted a preliminary study to evaluate the effectiveness of PLMs and state-of-the-art LLMs across seven popular programming languages. Our findings reveal that the PLM CodeT5P achieves the best performance in multilingual vulnerability detection, particularly in identifying the most critical vulnerabilities. Based on these results, we further discuss the potential of LLMs in advancing real-world multilingual vulnerability detection. This work represents an initial step toward exploring PLMs and LLMs for cross-language vulnerability detection, offering key insights for future research and practical deployment.

cs.SE↗

arXiv2025

Semantic Integrity Constraints: Declarative Guardrails for AI-Augmented Data Processing Systems

AI-augmented data processing systems (DPSs) integrate large language models (LLMs) into query pipelines, allowing powerful semantic operations on structured and unstructured data. However, the reliability (a.k.a. trust) of these systems is fundamentally challenged by the potential for LLMs to produce errors, limiting their adoption in critical domains. To help address this reliability bottleneck, we introduce semantic integrity constraints (SICs) -- a declarative abstraction for specifying and enforcing correctness conditions over LLM outputs in semantic queries. SICs generalize traditional database integrity constraints to semantic settings, supporting common types of constraints, such as grounding, soundness, and exclusion, with both reactive and proactive enforcement strategies. We argue that SICs provide a foundation for building reliable and auditable AI-augmented data systems. Specifically, we present a system design for integrating SICs into query planning and runtime execution and discuss its realization in AI-augmented DPSs. To guide and evaluate our vision, we outline several design goals -- covering criteria around expressiveness, runtime semantics, integration, performance, and enterprise-scale applicability -- and discuss how our framework addresses each, along with open research challenges.

cs.DB↗

arXiv2025

RAGVA: Engineering Retrieval Augmented Generation-based Virtual Assistants in Practice

Retrieval-augmented generation (RAG)-based applications are gaining prominence due to their ability to leverage large language models (LLMs). These systems excel at combining retrieval mechanisms with generative capabilities, resulting in more accurate, contextually relevant responses that enhance user experience. In particular, Transurban, a road operation company, is replacing its rule-based virtual assistant (VA) with a RAG-based VA (RAGVA) to offer more flexible customer interactions and support a wider range of scenarios. In this paper, drawing from the experience at Transurban, we present a comprehensive step-by-step guide for building a conversational application and how to engineer a RAGVA. These guides aim to serve as references for future researchers and practitioners. While the engineering processes for traditional software applications are well-established, the development and evaluation of RAG-based applications are still in their early stages, with numerous emerging challenges remaining uncharted. To address this gap, we conduct a focus group study with Transurban practitioners regarding developing and evaluating their RAGVA. We identified eight challenges encountered by the engineering team and proposed eight future directions that should be explored to advance the development of RAG-based applications. This study contributes to the foundational understanding of a RAG-based conversational application and the emerging AI software engineering challenges it presents.

cs.SE↗

arXiv2024

AI for DevSecOps: A Landscape and Future Opportunities

DevOps has emerged as one of the most rapidly evolving software development paradigms. With the growing concerns surrounding security in software systems, the DevSecOps paradigm has gained prominence, urging practitioners to incorporate security practices seamlessly into the DevOps workflow. However, integrating security into the DevOps workflow can impact agility and impede delivery speed. Recently, the advancement of artificial intelligence (AI) has revolutionized automation in various software domains, including software security. AI-driven security approaches, particularly those leveraging machine learning or deep learning, hold promise in automating security workflows. They reduce manual efforts, which can be integrated into DevOps to ensure uninterrupted delivery speed and align with the DevSecOps paradigm simultaneously. This paper seeks to contribute to the critical intersection of AI and DevSecOps by presenting a comprehensive landscape of AI-driven security techniques applicable to DevOps and identifying avenues for enhancing security, trust, and efficiency in software development processes. We analyzed 99 research papers spanning from 2017 to 2023. Specifically, we address two key research questions (RQs). In RQ1, we identified 12 security tasks associated with the DevSecOps process and reviewed existing AI-driven security approaches, the problems they addressed, and the 65 benchmarks used to evaluate those approaches. Drawing insights from our findings, in RQ2, we discussed state-of-the-art AI-driven security approaches, highlighted 15 challenges in existing research, and proposed 15 corresponding avenues for future opportunities.

cs.SE↗

arXiv2023

ChatGPT for Vulnerability Detection, Classification, and Repair: How Far Are We?

Large language models (LLMs) like ChatGPT (i.e., gpt-3.5-turbo and gpt-4) exhibited remarkable advancement in a range of software engineering tasks associated with source code such as code review and code generation. In this paper, we undertake a comprehensive study by instructing ChatGPT for four prevalent vulnerability tasks: function and line-level vulnerability prediction, vulnerability classification, severity estimation, and vulnerability repair. We compare ChatGPT with state-of-the-art language models designed for software vulnerability purposes. Through an empirical assessment employing extensive real-world datasets featuring over 190,000 C/C++ functions, we found that ChatGPT achieves limited performance, trailing behind other language models in vulnerability contexts by a significant margin. The experimental outcomes highlight the challenging nature of vulnerability prediction tasks, requiring domain-specific expertise. Despite ChatGPT's substantial model scale, exceeding that of source code-pre-trained language models (e.g., CodeBERT) by a factor of 14,000, the process of fine-tuning remains imperative for ChatGPT to generalize for vulnerability prediction tasks. We publish the studied dataset, experimental prompts for ChatGPT, and experimental results at https://github.com/awsm-research/ChatGPT4Vul.

cs.SE↗

arXiv2023

AIBugHunter: A Practical Tool for Predicting, Classifying and Repairing Software Vulnerabilities

Many ML-based approaches have been proposed to automatically detect, localize, and repair software vulnerabilities. While ML-based methods are more effective than program analysis-based vulnerability analysis tools, few have been integrated into modern IDEs, hindering practical adoption. To bridge this critical gap, we propose AIBugHunter, a novel ML-based software vulnerability analysis tool for C/C++ languages that is integrated into Visual Studio Code. AIBugHunter helps software developers to achieve real-time vulnerability detection, explanation, and repairs during programming. In particular, AIBugHunter scans through developers' source code to (1) locate vulnerabilities, (2) identify vulnerability types, (3) estimate vulnerability severity, and (4) suggest vulnerability repairs. In this article, we propose a novel multi-objective optimization (MOO)-based vulnerability classification approach and a transformer-based estimation approach to help AIBugHunter accurately identify vulnerability types and estimate severity. Our empirical experiments on a large dataset consisting of 188K+ C/C++ functions confirm that our proposed approaches are more accurate than other state-of-the-art baseline methods for vulnerability classification and estimation. Furthermore, we conduct qualitative evaluations including a survey study and a user study to obtain software practitioners' perceptions of our AIBugHunter tool and assess the impact that AIBugHunter may have on developers' productivity in security aspects. Our survey study shows that our AIBugHunter is perceived as useful where 90% of the participants consider adopting our AIBugHunter. Last but not least, our user study shows that our AIBugHunter could possibly enhance developers' productivity in combating cybersecurity issues during software development.

cs.SE↗