arXiv subjects

Zhenkai Liang

Publications and source records attributed to Zhenkai Liang.

At least 19 recordsLinked to original sources

arXiv2026

LMSM: LLM Security Framework Inspired by Linux Security Modules

Large language models (LLMs) are increasingly deployed with layered defenses, yet malicious prompts can still bypass them. Interpretability methods can expose model-internal signals along the generation path that could inform enforcement, but these signals are not security controls by themselves. Deployments that adapt them for safety typically couple each signal to its own calibration, policy logic, and intervention code, so each new artifact creates integration work instead of strengthening a shared defense. We present Language Model Security Modules (LMSM), a security framework that adapts the separation behind Linux Security Modules (LSM) to LLM serving. In LMSM, a selected security backend exposes calibrated evidence, a versioned policy evaluates active rules over trusted per-request context, and a separate gate authorizes buffered output release. This design separates mediation correctness from policy effectiveness, and it allows backend, rule, or schedule changes without rebuilding request handling or enforcement. Our prototype shows the separation working in practice: with Hugging Face Transformers and continuously batched vLLM, the same substrate hosts artifact-backed sparse autoencoder (SAE) and transcoder deployments and task-fitted dense probes, preserves request-specific decisions under scheduler churn, and selectively enforces and composes multiple rules per request. On Qwen3-4B, LMSM-Checkpoint reduces HarmBench attack success rate from 39.20% to 3.32%, with XSTest false refusals rising from 2.40% to 4.40%, while retaining 98.14% of the throughput of a matched serving path that performs no monitoring work at 32 active sequences. LMSM gives advances in interpretability and model-internal analysis a common path to runtime enforcement.

cs.CR↗

arXiv2026

SysEvolve: An AI-native, safe, autonomous adversarial attack-defense co-evolutionary system

The rapid advancement of large language models (LLMs) has created a growing asymmetry in cybersecurity, where attack accelerates toward autonomous execution while defense remains predominantly human-intensive. Despite substantial prior work across cyber ranges, AI-driven attack, and AI-driven defense, this asymmetry persists. We trace it to a deeper root cause, that evolution itself has stalled on both sides at three layers. To overcome this, we propose co-evolution as the integrating insight, where attack and defense AI agents autonomously and safely drive each other's evolution through adversarial confrontation. Based on this insight, we present \sysevolve, comprising three co-designed components, \sysfield, \sysspear, and \sysarmor. \sysfield constructs realistic multi-host ranges. \sysspear generates efficient, safe attack schemes. \sysarmor performs real-time, interpretable defense. Together they form a self-driven adversarial loop restoring evolution at all three layers. In evaluation, \sysfield achieves zero-loss collection at 2.1\% overhead and orchestrates 257 CVEs into 1,148 ranges, \sysspear improves attack success by over 25\% over baseline LLMs, and \sysarmor achieves 10--1000$\times$ greater precision than prior systems and detects real APT attacks in production at Huawei and Sangfor. Our evaluation also reveals three findings about LLM agent capabilities. First, multi-step composition and larger topologies expose agent capability gaps hidden by single-step evaluations. Second, the bottleneck lies after initial access in post-compromise state utilization. Third, LLM agents are susceptible to environmental interference. When decoy endpoints are deployed in the range, agent timeouts triple and downstream completion disappears despite the success rates of initial accesses are unchanged.

cs.CR↗

arXiv2026

Cosine Misleads: Auxiliary Losses Reshape Vision Language Models, Not Their Latents

Latent visual reasoning (LVR) inserts supervised latent tokens between perception and answer generation in vision-language models (VLMs). The field uses alignment between these latents and their visual targets, i.e., cosine similarity or mean squared error (MSE), as both the training loss and the quality metric, assuming that better alignment yields a better answer. We test this with a designed matrix of five LVR variants and find the assumption inverted: cosine alignment is negatively correlated with accuracy across all five (r=-0.94). To explain this, we introduce PRISM, a pair of inference-time diagnostics: a linear probe that asks where the answer is decodable, and a corruption test that asks whether the latent is load-bearing. The supervised latents are largely bypassed. Corrupting them shifts accuracy by at most four points. The answer is decodable downstream of the latent but not at it, and the size of this decodability gap predicts how much each variant relies on its latent under perturbation. Consistent with an Information Bottleneck reading of the loss, the auxiliary objective reshapes the language model via shared parameters rather than via the latent variable it nominally optimizes.

cs.CV↗

arXiv2026

Self-Evaluation Is Already There: Eliciting Latent Judge Calibration in Base LLMs with Minimal Data

Large language models are increasingly evaluated by other models, raising a natural question: can a model predict how a judge will score its own output? We find that the ability is largely present before any targeted training: prompted few-shot, a base model already predicts an external judge's multi-attribute quality scores on open-ended responses well above chance across three benchmarks. We introduce Self-Evaluation Elicitation (SEE), a method that surfaces this latent ability through a short cycle comprising a calibration-coupled reinforcement learning phase that improves the answer and predicts the judge, followed by a masked distillation phase that sharpens the prediction while leaving the answer untouched. From 160 unique examples, roughly 31x fewer than a reinforcement learning baseline, SEE improves held-out calibration across three benchmarks while preserving answer quality. The elicited self-evaluation is sharply localized within the model's own token distribution and stable across judges it was never trained against, indicating a transferable notion of quality rather than a single judge's preference. These results reframe judge-aligned self-evaluation as a problem of elicitation rather than acquisition.

cs.CL↗

arXiv2026

Heimdallr: Characterizing and Detecting LLM-Induced Security Risks in GitHub CI Workflows

GitHub Continuous Integration (CI) workflows increasingly integrate Large Language Models (LLMs) to automate review, triage, content generation, and repository maintenance. This creates a new attack surface: externally controllable workflow inputs can shape LLM prompts and outputs, which may in turn affect security decisions, repository state, or privileged execution. Although LLM security and CI security have each been studied extensively, their intersection remains underexplored. In this paper, we present the first study of LLM-induced security risks in GitHub CI workflows. We characterize the problem along the full execution chain and develop a taxonomy of high-level risk classes and concrete threat vectors. To detect such risks in practice, we design Heimdallr, a hybrid analysis framework that normalizes workflows into an LLM-Workflow Property Graph (L-WPG) and combines triggerability analysis, LLM-assisted dataflow summarization, and deterministic propagation to synthesize concrete threat-vector findings. Evaluated on 300 manually annotated unique workflows, Heimdallr achieves high accuracy on LLM-node identification (F1~=~0.994), triggerability classification (99.8%), and threat-vector detection (micro-average F1~=~0.917). As part of an ongoing detection and disclosure effort, we have so far responsibly disclosed 802 vulnerable workflow instances across 759 repositories and received 71 acknowledgments.

cs.CR↗

arXiv2026

Unraveling the Key of Machine Learning-based Android Malware Detection

With the rapid advancement of machine learning (ML), ML-based Android malware detection has gained significant popularity due to its ability to automatically learn malicious patterns from Android apps. However, the lack of an in-depth and systematic analysis of existing research makes it difficult to obtain a holistic understanding of the state of the art in this field. In this work, we present the most comprehensive investigation to date of ML-based Android malware detection systems, combining both empirical and quantitative analyses. We first organize prior work into a unified taxonomy based on Android app representations and the ML modeling pipeline. Building on this taxonomy, we design a general-purpose framework for ML-based Android malware detection and re-implement 12 representative approaches from three research communities -- software engineering, security, and machine learning. Using this framework, we conduct a large-scale evaluation across three key dimensions: detection effectiveness, robustness to real-world challenges, and efficiency. Despite extensive research efforts and encouraging results, our findings reveal that existing learning-based Android malware detectors still face significant challenges, including vulnerability to malware evolution and susceptibility to adversarial attacks. We attribute these limitations to the detectors' ability to capture and leverage malware semantics, defined as semantic information that characterizes malicious behaviors derived from APK features. Finally, we summarize our key insights and provide actionable recommendations to guide future research in this domain.

cs.CR↗

arXiv2026

AttackSeqBench: Benchmarking the Capabilities of LLMs for Attack Sequences Understanding

Cyber Threat Intelligence (CTI) reports document observations of cyber threats, synthesizing evidence about adversaries' actions and intent into actionable knowledge that informs detection, response, and defense planning. However, the unstructured and verbose nature of CTI reports poses significant challenges for security practitioners to manually extract and analyze such sequences. Although large language models (LLMs) exhibit promise in cybersecurity tasks such as entity extraction and knowledge graph construction, their understanding and reasoning capabilities towards behavioral sequences remains underexplored. To address this, we introduce AttackSeqBench, a benchmark designed to systematically evaluate LLMs' reasoning abilities across the tactical, technical, and procedural dimensions of adversarial behaviors, while satisfying Extensibility, Reasoning Scalability, and Domain-dpecific Epistemic Expandability. We further benchmark 7 LLMs, 5 LRMs and 4 post-training strategies across 3 benchmark settings and 3 benchmark tasks within our AttackSeqBench to identify their advantages and limitations in such specific domain. Our findings contribute to a deeper understanding of LLM-driven CTI report understanding and foster its application in cybersecurity operations. Our code of benchmark construction and evaluation and the corresponding dataset are available at: https://github.com/hulkima/AttackSeqBench.

cs.CR↗

arXiv2026

Do LLMs and VLMs Share Neurons for Inference? Evidence and Mechanisms of Cross-Modal Transfer

Large vision-language models (LVLMs) have rapidly advanced across various domains, yet they still lag behind strong text-only large language models (LLMs) on tasks that require multi-step inference and compositional decision-making. Motivated by their shared transformer architectures, we investigate whether the two model families rely on common internal computation for such inference. At the neuron level, we uncover a surprisingly large overlap: more than half of the top-activated units during multi-step inference are shared between representative LLMs and LVLMs, revealing a modality-invariant inference subspace. Through causal probing via activation amplification, we further show that these shared neurons encode consistent and interpretable concept-level effects, demonstrating their functional contribution to inference. Building on this insight, we propose Shared Neuron Low-Rank Fusion (SNRF), a parameter-efficient framework that transfers mature inference circuitry from LLMs to LVLMs. SNRF profiles cross-model activations to identify shared neurons, computes a low-rank approximation of inter-model weight differences, and injects these updates selectively within the shared-neuron subspace. This mechanism strengthens multimodal inference performance with minimal parameter changes and requires no large-scale multimodal fine-tuning. Across diverse mathematics and perception benchmarks, SNRF consistently enhances LVLM inference performance while preserving perceptual capabilities. Our results demonstrate that shared neurons form an interpretable bridge between LLMs and LVLMs, enabling low-cost transfer of inference ability into multimodal models. Our code is available at [https://github.com/chenhangcuisg-code/Do-LLMs-VLMs-Share-Neurons](https://github.com/chenhangcuisg-code/Do-LLMs-VLMs-Share-Neurons).

cs.CL↗

arXiv2026

AlphaSteer: Learning Refusal Steering with Principled Null-Space Constraint

As LLMs are increasingly deployed in real-world applications, ensuring their ability to refuse malicious prompts, especially jailbreak attacks, is essential for safe and reliable use. Recently, activation steering has emerged as an effective approach for enhancing LLM safety by adding a refusal direction vector to internal activations of LLMs during inference, which will further induce the refusal behaviors of LLMs. However, indiscriminately applying activation steering fundamentally suffers from the trade-off between safety and utility, since the same steering vector can also lead to over-refusal and degraded performance on benign prompts. Although prior efforts, such as vector calibration and conditional steering, have attempted to mitigate this trade-off, their lack of theoretical grounding limits their robustness and effectiveness. To better address the trade-off between safety and utility, we present a theoretically grounded and empirically effective activation steering method called AlphaSteer. Specifically, it considers activation steering as a learnable process with two principled learning objectives: utility preservation and safety enhancement. For utility preservation, it learns to construct a nearly zero vector for steering benign data, with the null-space constraints. For safety enhancement, it learns to construct a refusal direction vector for steering malicious data, with the help of linear regression. Experiments across multiple jailbreak attacks and utility benchmarks demonstrate the effectiveness of AlphaSteer, which significantly improves the safety of LLMs without compromising general capabilities. Our codes are available at https://github.com/AlphaLab-USTC/AlphaSteer.

cs.LG↗

arXiv2026

Self-Guard: Defending Large Reasoning Models via enhanced self-reflection

The emergence of Large Reasoning Models (LRMs) introduces a new paradigm of explicit reasoning, enabling remarkable advances yet posing unique risks such as reasoning manipulation and information leakage. To mitigate these risks, current alignment strategies predominantly rely on heavy post-training paradigms or external interventions. However, these approaches are often computationally intensive and fail to address the inherent awareness-compliance gap, a critical misalignment where models recognize potential risks yet prioritize following user instructions due to their sycophantic tendencies. To address these limitations, we propose Self-Guard, a lightweight safety defense framework that reinforces safety compliance at the representational level. Self-Guard operates through two principal stages: (1) safety-oriented prompting, which activates the model's latent safety awareness to evoke spontaneous reflection, and (2) safety activation steering, which extracts the resulting directional shift in the hidden state space and amplifies it to ensure that safety compliance prevails over sycophancy during inference. Experiments demonstrate that Self-Guard effectively bridges the awareness-compliance gap, achieving robust safety performance without compromising model utility. Furthermore, Self-Guard exhibits strong generalization across diverse unseen risks and varying model scales, offering a cost-efficient solution for LRM safety alignment.

cs.AI↗

arXiv2026

DevOps-Gym: Benchmarking AI Agents in Software DevOps Cycle

Even though demonstrating extraordinary capabilities in code generation and software issue resolving, AI agents' capabilities in the full software DevOps cycle are still unknown. Different from pure code generation, handling the DevOps cycle in real-world software, including developing, deploying, and managing, requires analyzing large-scale projects, understanding dynamic program behaviors, leveraging domain-specific tools, and making sequential decisions. However, existing benchmarks focus on isolated problems and lack environments and tool interfaces for DevOps. We introduce DevOps-Gym, the first end-to-end benchmark for evaluating AI agents across core DevOps workflows: build and configuration, monitoring, issue resolving, and test generation. DevOps-Gym includes 700+ real-world tasks collected from 30+ projects in Java and Go. We develop a semi-automated data collection mechanism with rigorous and non-trivial expert efforts in ensuring the task coverage and quality. Our evaluation of state-of-the-art models and agents reveals fundamental limitations: they struggle with issue resolving and test generation in Java and Go, and remain unable to handle new tasks such as monitoring and build and configuration. These results highlight the need for essential research in automating the full DevOps cycle with AI agents.

cs.SE↗

arXiv2025

RSafe: Incentivizing proactive reasoning to build robust and adaptive LLM safeguards

Large Language Models (LLMs) continue to exhibit vulnerabilities despite deliberate safety alignment efforts, posing significant risks to users and society. To safeguard against the risk of policy-violating content, system-level moderation via external guard models-designed to monitor LLM inputs and outputs and block potentially harmful content-has emerged as a prevalent mitigation strategy. Existing approaches of training guard models rely heavily on extensive human curated datasets and struggle with out-of-distribution threats, such as emerging harmful categories or jailbreak attacks. To address these limitations, we propose RSafe, an adaptive reasoning-based safeguard that conducts guided safety reasoning to provide robust protection within the scope of specified safety policies. RSafe operates in two stages: 1) guided reasoning, where it analyzes safety risks of input content through policy-guided step-by-step reasoning, and 2) reinforced alignment, where rule-based RL optimizes its reasoning paths to align with accurate safety prediction. This two-stage training paradigm enables RSafe to internalize safety principles to generalize safety protection capability over unseen or adversarial safety violation scenarios. During inference, RSafe accepts user-specified safety policies to provide enhanced safeguards tailored to specific safety requirements.

cs.AI↗

arXiv2025

PromoGuardian: Detecting Promotion Abuse Fraud with Multi-Relation Fused Graph Neural Networks

As e-commerce platforms develop, fraudulent activities are increasingly emerging, posing significant threats to the security and stability of these platforms. Promotion abuse is one of the fastest-growing types of fraud in recent years and is characterized by users exploiting promotional activities to gain financial benefits from the platform. To investigate this issue, we conduct the first study on promotion abuse fraud in e-commerce platforms MEITUAN. We find that promotion abuse fraud is a group-based fraudulent activity with two types of fraudulent activities: Stocking Up and Cashback Abuse. Unlike traditional fraudulent activities such as fake reviews, promotion abuse fraud typically involves ordinary customers conducting legitimate transactions and these two types of fraudulent activities are often intertwined. To address this issue, we propose leveraging additional information from the spatial and temporal perspectives to detect promotion abuse fraud. In this paper, we introduce PROMOGUARDIAN, a novel multi-relation fused graph neural network that integrates the spatial and temporal information of transaction data into a homogeneous graph to detect promotion abuse fraud. We conduct extensive experiments on real-world data from MEITUAN, and the results demonstrate that our proposed model outperforms state-of-the-art methods in promotion abuse fraud detection, achieving 93.15% precision, detecting 2.1 to 5.0 times more fraudsters, and preventing 1.5 to 8.8 times more financial losses in production environments.

cs.CR↗

arXiv2025

TraceAegis: Securing LLM-Based Agents via Hierarchical and Behavioral Anomaly Detection

LLM-based agents have demonstrated promising adaptability in real-world applications. However, these agents remain vulnerable to a wide range of attacks, such as tool poisoning and malicious instructions, that compromise their execution flow and can lead to serious consequences like data breaches and financial loss. Existing studies typically attempt to mitigate such anomalies by predefining specific rules and enforcing them at runtime to enhance safety. Yet, designing comprehensive rules is difficult, requiring extensive manual effort and still leaving gaps that result in false negatives. As agent systems evolve into complex software systems, we take inspiration from software system security and propose TraceAegis, a provenance-based analysis framework that leverages agent execution traces to detect potential anomalies. In particular, TraceAegis constructs a hierarchical structure to abstract stable execution units that characterize normal agent behaviors. These units are then summarized into constrained behavioral rules that specify the conditions necessary to complete a task. By validating execution traces against both hierarchical and behavioral constraints, TraceAegis is able to effectively detect abnormal behaviors. To evaluate the effectiveness of TraceAegis, we introduce TraceAegis-Bench, a dataset covering two representative scenarios: healthcare and corporate procurement. Each scenario includes 1,300 benign behaviors and 300 abnormal behaviors, where the anomalies either violate the agent's execution order or break the semantic consistency of its execution sequence. Experimental results demonstrate that TraceAegis achieves strong performance on TraceAegis-Bench, successfully identifying the majority of abnormal behaviors.

cs.CR↗

arXiv2025

Propagation-Based Vulnerability Impact Assessment for Software Supply Chains

Identifying the impact scope and scale is critical for software supply chain vulnerability assessment. However, existing studies face substantial limitations. First, prior studies either work at coarse package-level granularity, producing many false positives, or fail to accomplish whole-ecosystem vulnerability propagation analysis. Second, although vulnerability assessment indicators like CVSS characterize individual vulnerabilities, no metric exists to specifically quantify the dynamic impact of vulnerability propagation across software supply chains. To address these limitations and enable accurate and comprehensive vulnerability impact assessment, we propose a novel approach: (i) a hierarchical worklist-based algorithm for whole-ecosystem and call-graph-level vulnerability propagation analysis and (ii) the Vulnerability Propagation Scoring System (VPSS), a dynamic metric to quantify the scope and evolution of vulnerability impacts in software supply chains. We implement a prototype of our approach in the Java Maven ecosystem and evaluate it on 100 real-world vulnerabilities. Experimental results demonstrate that our approach enables effective ecosystem-wide vulnerability propagation analysis, and provides a practical, quantitative measure of vulnerability impact through VPSS.

cs.SE↗

arXiv2025

When MCP Servers Attack: Taxonomy, Feasibility, and Mitigation

Model Context Protocol (MCP) servers enable AI applications to connect to external systems in a plug-and-play manner, but their rapid proliferation also introduces severe security risks. Unlike mature software ecosystems with rigorous vetting, MCP servers still lack standardized review mechanisms, giving adversaries opportunities to distribute malicious implementations. Despite this pressing risk, the security implications of MCP servers remain underexplored. To address this gap, we present the first systematic study that treats MCP servers as active threat actors and decomposes them into core components to examine how adversarial developers can implant malicious intent. Specifically, we investigate three research questions: (i) what types of attacks malicious MCP servers can launch, (ii) how vulnerable MCP hosts and Large Language Models (LLMs) are to these attacks, and (iii) how feasible it is to carry out MCP server attacks in practice. Our study proposes a component-based taxonomy comprising twelve attack categories. For each category, we develop Proof-of-Concept (PoC) servers and demonstrate their effectiveness across diverse real-world host-LLM settings. We further show that attackers can generate large numbers of malicious servers at virtually no cost. We then test state-of-the-art scanners on the generated servers and found that existing detection approaches are insufficient. These findings highlight that malicious MCP servers are easy to implement, difficult to detect with current tools, and capable of causing concrete damage to AI agent systems. Addressing this threat requires coordinated efforts among protocol designers, host developers, LLM providers, and end users to build a more secure and resilient MCP ecosystem.

cs.CR↗

arXiv2025

PsyScam: A Benchmark for Psychological Techniques in Real-World Scams

Over the years, online scams have grown dramatically, with nearly 50% of global consumers encountering scam attempts each week. These scams cause not only significant financial losses to individuals and businesses, but also lasting psychological trauma, largely due to scammers' strategic employment of psychological techniques (PTs) to manipulate victims. Meanwhile, scammers continually evolve their tactics by leveraging advances in Large Language Models (LLMs) to generate diverse scam variants that easily bypass existing defenses. To address this pressing problem, we introduce PsyScam, a benchmark designed to systematically capture the PTs employed in real-world scam reports, and investigate how LLMs can be utilized to generate variants of scams based on the PTs and the contexts provided by these scams. Specifically, we collect a wide range of scam reports and ground its annotations of employed PTs in well-established cognitive and psychological theories. We further demonstrate LLMs' capabilities in generating through two downstream tasks: scam completion, and scam augmentation. Experimental results show that PsyScam presents significant challenges to existing models in both detecting and generating scam content based on the PTs used by real-world scammers. Our code and dataset are available.

cs.CR↗

arXiv2025

PiMRef: Deducing Ever-evolving Spear-phishing Emails with Knowledge Base Invariants

Phishing email is a critical step in the cybercrime kill chain due to the high reachability of victims' email accounts and the low cost of launching phishing campaigns. This ever-evolving nature of phishing emails makes traditional rule-based and feature-engineering-based phishing email detectors fight an uphill battle in the cat-and-mouse game of defense and attack. In this work, we show that, large language models (LLMs) can be effectively exploited to generate profile-grounded spear-phishing, compromising major paradigms of phishing email detectors. To defend against such LLM-based spear-phishing attacks, we propose PiMRef, the first reference-based solution to detect ever-evolving phishing emails using knowledge-based invariants, targeting the identity-impersonation attacks that characterize spear-phishing. Our rationale lies in the fact that convincing phishing emails often include ``disprovable claims'', which contradict certain real-world facts. Technically, given an email, PiMRef (i) discovers the claimed identity of the sender, (ii) verifies the sender's email domain against a dynamically expandable knowledge base, and (iii) infers call-to-action instructions that encourage next-step engagement. Compared to existing baselines such as D-Fence, HelpHed, and ChatSpamDetector, PimRef reduces the false-positive rate to 0.81\% while maintaining a recall of 90.7\%-93.1\% on conventional phishing benchmarks such as Nazario and PhishPot. On SpearMail, our newly constructed benchmark of 14,672 LLM-generated spear-phishing emails targeting 681 public profiles, PimRef reaches a recall of 86.4\% without incurring additional false positives. Our code is publicly available at https://github.com/code-philia/PhishEmail.

cs.CR↗