defense implementation

In threat hunting, analysts struggle to identify co-occurring Tactics, Techniques, and Procedures (TTPs) from massive network traffic and fragmented cyber threat intelligence (CTI). To address this, we propose the first implicit-feedback recommendation framework for TTP association inference. Our method constructs the largest ATT&CK-aligned CTI report annotation dataset to date; introduces Bayesian Personalized Ranking (BPR) and LightGCN—state-of-the-art implicit-feedback models—into TTP co-occurrence modeling to mitigate intelligence omission; and integrates t-SNE visualization with interpretable embedding spaces to enable tactic-level association analysis. Experiments demonstrate significant improvements in TTP association prediction accuracy. All code, data, and an interactive web interface are publicly released to support operational threat hunting.

This work proposes a dynamic attack path generation framework that integrates large language models (LLMs), white-box system access, and Monte Carlo Tree Search (MCTS) to automatically construct and validate attack paths in real-world vulnerable environments without relying on pre-built vulnerability graphs or exploit databases. Unlike traditional cyber defense exercises that depend on manual path construction by experts, the proposed approach achieves end-to-end attack path generation without prior knowledge of the target environment, thereby shifting the expert’s role from technical validation to scenario design and significantly accelerating development cycles. Evaluated in the CIDeX 2025 exercise involving 46 hosts, the generated paths matched human-crafted scenarios in terms of pedagogical value, engagement, credibility, and challenge level, while reducing scenario development time from months to days.

This study addresses the limitations of existing threat hunting tools in supporting cognitive externalization, clue organization, and cross-session continuity, which hinder effective collaboration and reasoning among threat hunters. Building upon prior research on mental model construction and sharing, the authors propose six design heuristics to evaluate the cognitive support capabilities of threat hunting tools. Guided by these heuristics, they developed and implemented a prototype system, Threat Hunter Board. Employing the design science research paradigm and cognitive walkthroughs, the development and preliminary evaluation demonstrate the tool’s feasibility in enhancing cognitive and collaborative workflows. This work lays the groundwork for future empirical studies with professional users.

This work addresses the precision-recall trade-off in network intrusion detection, arising from the inherent diversity of cyber threats. Methodologically, we propose a novel detection framework that deeply integrates a cybersecurity incident taxonomy into the architectural design of detection networks. Guided by the taxonomy’s hierarchical semantic structure, we jointly leverage ontology-based analysis and controlled simulations to systematically identify the optimal operational equilibrium for detection strategies. Our key contribution is the first principled elevation of taxonomies from static labeling tools to structural priors embedded within detection models—explicitly encoding semantic relationships and evolutionary pathways among threat behaviors. Empirical evaluation across multiple public benchmark datasets demonstrates substantial improvements in holistic detection performance (average F1-score gain of 12.7%). Moreover, our analysis uncovers fundamental theoretical limits on detection set construction and establishes an interpretable pathway for performance optimization.

AI systems often exhibit harmful behaviors due to unknown failure modes, yet conventional red-teaming and adversarial training rely on known attack samples, limiting generalization to unseen threats. Method: We propose Latent-space Adversarial Training (LAT), the first approach to inject adversarial perturbations directly into structured latent representations—without requiring prior knowledge of attack patterns or concrete adversarial examples. LAT integrates latent-space modeling, adversarial perturbation of latent variables, and multi-task joint optimization, and is applicable to image classification, text classification, and text generation. Contribution/Results: Experiments demonstrate that LAT significantly improves robustness against unseen attack types, backdoor triggers, and out-of-distribution adversarial examples, while preserving or even enhancing clean-sample accuracy. By decoupling adversarial defense from reliance on known attack templates, LAT establishes a general-purpose defense framework for previously unobserved failure modes.

This study addresses the challenge faced by resource-constrained organizations in translating cybersecurity governance frameworks—such as the NIST Cybersecurity Framework (CSF)—into actionable defense decisions. The authors propose an integrated optimization approach that synergistically combines governance requirements, adversary knowledge from MITRE ATT&CK, and adversarially aware learning. By establishing a mapping between NIST CSF maturity levels and ATT&CK mitigations, modeling attack paths via a variable-order Markov model, and formulating a deep reinforcement learning framework, the method generates cost-effective and risk-balanced defense strategies under budget constraints. This work represents the first effort to cohesively unify governance frameworks, ATT&CK-based threat intelligence, and interpretable reinforcement learning, yielding operationally viable mitigation plans that align with an organization’s security maturity and exhibit robustness against adaptive adversaries.

This work proposes a general-purpose red-teaming framework that overcomes the limitations of existing automated approaches, which are often confined to specific security scenarios and rely on evaluators known during training, thereby lacking generalization to novel adversarial targets. By end-to-end fine-tuning compact language models such as Qwen3-8B and integrating multi-objective adversarial example generation with adaptive optimization strategies, the method generates effective attacks against arbitrary red-teaming tasks without requiring predefined evaluators. Experimental results demonstrate significant improvements in attack generation performance both within and across domains. To the best of our knowledge, this is the first approach to achieve evaluator-agnostic, generalizable red-teaming automation, effectively transcending the constraints of conventional methods in terms of task scope and adaptability.

This study addresses the absence of a unified safety evaluation benchmark for current large language models (LLMs), which hinders the quantification of deployment risks in critical applications. To bridge this gap, we propose the first standardized, cross-architecture safety evaluation framework that systematically assesses the vulnerability of five representative LLMs under six categories of adversarial attacks. Furthermore, we introduce a deployable multi-layered external defense mechanism. Experimental results reveal that existing models exhibit vulnerability rates ranging from 11.9% to 29.8%, whereas our defense framework achieves an average detection accuracy of 83% with only a 5% false positive rate. Notably, the findings demonstrate no direct correlation between a model’s general capabilities and its safety robustness, offering empirical evidence and a practical solution for safer LLM deployment.

This work addresses the limited defensive capabilities in cyber-physical systems by proposing a multilayer collaborative deception strategy. The approach models attacker-defender interactions as a Stackelberg game and leverages real-world vulnerability data from the National Vulnerability Database (NVD) along with CVSS scores to synchronously deploy deception nodes across both cyber and physical domains, thereby disrupting adversarial reconnaissance. A novel CVE-based utility function is designed to identify critical vulnerabilities and enable coordinated optimization of deception resources across domains. Experimental results demonstrate that the proposed method significantly enhances defender utility under both CVSS v2 and v3 metrics, outperforming single-layer and baseline defense schemes, and effectively overcomes the limitations of traditional single-domain defense strategies.

This study investigates human users’ perceptual vulnerability to Agent-Mediated Deception (AMD) attacks in large language model (LLM)-driven agent systems. Leveraging a high-fidelity experimental platform, HAT-Lab, the authors conducted a large-scale empirical study with 303 participants across nine scenarios spanning everyday and professional contexts. The research reveals, for the first time, critical cognitive failure modes among users confronting compromised agents: only 8.6% detected the attack, and experts were paradoxically more susceptible in certain scenarios. To address this vulnerability, the work proposes a novel defense mechanism centered on low-cost verification, workflow-interruptive warnings, and experiential learning. Experimental results demonstrate that over 90% of users significantly enhanced their vigilance after experiential training, confirming the effectiveness and feasibility of the proposed defensive approach.