The new-round scientific and technological revolution and industrial transformation are reshaping the global landscape at an unprecedented pace. As a strategic technology driving this transformation, artificial intelligence has become a core focus of global technological competition. Against the backdrop of advancing the “AI+” initiative and fostering new-quality productive forces, artificial intelligence in China has pervaded science and technology, industry, people’s livelihood, governance and other fields. While unlocking socio-economic momentum, it has also given rise to multiple security-related challenges including supply-chain risks, algorithmic ethics, employment polarization. Balancing development and security has gradually become the central thread of artificial intelligence governance. At present, domestic practices of AI integration tend to prioritize industrial scale expansion with insufficient consideration of potential risk costs. Some applications remain at the shallow implementation stage, and in-depth integration needs further improvement. Evaluation tools that take both development benefits and security risks into account are still inadequate. This paper sorts out the multi-dimensional impacts of AI-integrated development on economy, industry, employment and global competition, and analyzes the practical bottlenecks in evaluating its effectiveness. Drawing on differentiated evaluation approaches from the United States, the European Union, Japan and South Korea, it puts forward suggestions on evaluation pathways, so as to provide references for judging the effectiveness of AI-integrated development and coordinating innovative development with whole?chain risk prevention and control.

The rapid development of Artificial Intelligence (AI) technologies and the expansion of their applications place greater demands on the development and effective implementation of AI safety and security standards. This paper reviews the AI safety and security standardization initiatives of international organizations, including ISO/IEC, ITU, and IEEE, compares standardization practices in the United States, the European Union, and China, and analyzes the major challenges facing China in this field. It argues that AI safety and security standards are evolving from individual technical specifications toward lifecycle-wide governance. However, China continues to face several challenges: standards development lags behind rapid technological advances; general safety and security requirements are difficult to reconcile fully with sector-specific needs; and international differences in governance approaches complicate standards alignment. To address these challenges, this paper recommends strengthening research on key technologies and coordination between research and development, testing, and evaluation; improving domestic coordination and alignment between general standards and sector-specific specifications; and advancing alignment with international rules and joint validation. These measures aim to enhance the verifiability, adaptability, and effectiveness of China’s AI safety and security standards and strengthen its capacity to participate in international standardization.

Large model security is a core issue in Artificial Intelligence(AI) governance, and its complexity far exceeds that of traditional cybersecurity and software vulnerability management. From the perspective of complex adaptive systems, this paper deconstructed large model security into a dynamic system shaped by the interaction of attack, defense, and vulnerability, examined its connotations, extensions, and constituent elements, and identified three major trends in its evolution. It argued that attack was essentially adaptive perturbation, defense was essentially resilience enhancing, and vulnerability was a tension node in the system structure that embodied both destructive potential and improvement value. The security boundary of large models is expanding from discrete points and chains toward the full life cycle, with technology, institutions, and openness jointly sustaining its effective operation. Accordingly, this paper put forward recommendations across three dimensions, namely technological offense and defense, governance rules, and international cooperation, so as to promote the construction of a security theory system commensurate with AI development and better implement the strategic requirement of coordinating development and security.

With the rapid proliferation of generative Artificial Intelligence (AI), large-scale models, and intelligent-agent applications, AI security risks have permeated data processing, model development, product deployment, and infrastructure operation. Accordingly, this study analyzed the characteristics and implementation approaches of AI safety governance in major countries and regions, and found that although the choices of governance measures vary significantly due to differences in technological foundations, legal traditions and security environments, convergent practices exist in fields including data privacy, algorithmic models, generative content, tiered regulation and technology supply security. This comparative study of international practices can provide valuable insights for China to refine its artificial intelligence safety governance system.

The explosive advancement of artificial intelligence is fundamentally reshaping the circulation patterns and security boundaries of data elements. As a new type of production factor, data has become the core engine driving new quality productive forces and the digital economy, with circulation security serving as a critical prerequisite for unlocking its value. Based on the characteristics of data elements and the implications of secure data circulation in the age of Artificial Intelligence(AI), this paper provides a systematic review of strategic orientations and institutional practices in the domain of global data circulation security. It further analyzes China’s current developmental status from multiple dimensions, including policies and regulations, technological support, and cross-border governance. The study reveals that, although China has preliminarily established a basic framework for data element circulation security, it still faces multiple structural challenges, such as an incomplete governance system, heavy reliance on core foreign technologies, insufficient alignment with international rules, and a shortage of interdisciplinary talent. To address these issues, this paper proposes a set of coordinated strategies: refining the ownership confirmation mechanism and standard system, strengthening R&D in key technologies, optimizing cross-border security management, and building collaborative talent cultivation mechanisms. These measures aim to advance both the governance system and capacity building, thereby providing theoretical references and practical pathways for developing a secure, trustworthy, open, and controllable data element market.

With the iterative evolution of generative artificial intelligence, large-scale models represented by DeepSeek are expected to shake off the past path dependence of “algorithm stacking”. Nevertheless, attendant risks in personal information processing keep emerging and are deeply embedded in the risk chain covering data input, training and output. Driven by the profit-seeking attributes of capital, such risks are superimposed, giving rise to multiple hidden dangers: in the horizontal dimension, the “notice-and-consent” rules become a mere formality; in the vertical dimension, personal information is on the verge of leakage; and in the systemic dimension, the foundation of social trust gradually weakens. These phenomena suggest an urgent need to revisit compliance governance strategies. In view of the above, at the horizontal dimension, the substantivization of rules shall be promoted to consolidate the compliance foundation, and mechanisms such as technical empowerment, regulatory constraints, traceability, and accountability shall be relied on to break the vicious cycle of risks. At the vertical dimension, rationalized privacy-protection mechanisms shall be adopted to adapt to the characteristics of information flow, so as to safeguard personal privacy while taking public interests into account. At the systemic dimension, standardized risk-prevention mechanisms shall be established to fulfill the bottom-line requirements of enterprise compliance and respond to the interests of diverse stakeholders. Based on the above-mentioned rule-of-law approaches, feasible governance solutions can be offered to mitigate personal-information-processing risks arising from generative artificial intelligence.

With the improvement of the cognitive ability of the large-scale model integration architecture represented by GPT-5, the automatic decision-making mode of administrative algorithms has also changed in many aspects, forming a paradigm transition from “rule driven” to “cognitive generalization”. The technical transition of administrative algorithm decision driven by the large-scale model integration architecture is specifically manifested in the cross-modal data integration, dynamic routing calculation force distribution and chain reasoning correction, which lays a new foundation for the automatic decision-making of administrative algorithms and improves the government’s digital governance ability and problem-solving ability. However at the same time, the deep involvement of the large-scale model integration architecture in the decision-making of administrative algorithms has also induced risks and challenges in data security, procedural justice and power operation structure to a certain extent. In view of this, corresponding institutional responses and regulations shall be formulated. On the one hand, it is necessary to reconstruct data governance boundaries through hierarchical data right confirmation and the public data trust system, and enhance the interpretability and transparency of algorithms via a dynamic interpretation mechanism. On the other hand, the principles of proportionality and anti-discrimination rules are embedded in the process of algorithm design, and the binding system of subject responsibility for algorithm decision-making is established.

The in-depth application of generative artificial intelligence has endowed intellectual property (IP) risks associated with large models with new notable characteristics, namely sudden onset, high transmissibility and severe destructiveness, which have posed substantial challenges to conventional governance frameworks. This study focuses on the IP emergencies triggered by the misuse of large model technologies, and aims to construct a response and disposal mechanism integrating intelligent early warning, rapid handling and collaborative coordination. First, this study establishes a multidimensional monitoring indicator system, multi-source data perception framework and dynamic grading rules to achieve precise risk identification and tiered classification. Subsequently, this study designs a standardized disposal workflow centered on immediate containment, evidence preservation and blockchain-based forensic storage, which ensures rapid interruption of risk propagation and provides solid support for subsequent legal accountability. Finally, this study develops a multi-stakeholder collaborative linkage mechanism covering government-enterprise cooperation, cross-sectoral coordination and cross-border interaction, so as to integrate resources, resolve jurisdictional conflicts and generate synergistic effects in risk prevention and control. The proposed mechanism aims to realize a paradigm shift from passive remediation to proactive prevention, and provides a theoretical framework and practical pathways for building an agile and efficient governance system for IP risks in the context of large models.

To address the privacy leakage risks existing in the application of Internet of Vehicles (IoV) devices, this paper systematically summarizes 359 types of IoV device privacy information classified into four categories by analyzing relevant IoV technical documents and academic papers, and constructs a corresponding privacy information classification system. Based on this classification system, this paper proposes a large language model (LLM)-based Vehicular Data-flow Privacy Compliance Analyzer (VDPAnalyzer) to conduct privacy compliance analysis on vehicular data streams. Based on the syntactic and semantic parsing of vehicular network traffic, the system integrates pattern matching and LLM technologies with prompt engineering and model fine-tuning strategies, so as to effectively detect privacy leakage problems and identify compliance risks in vehicular data streams. In this paper, network traffic data of 11 real-world IoV devices including on-board diagnostics (OBD) devices, vehicle-mounted cameras and in-vehicle infotainment systems are collected for experimental verification. The experimental results show that the fine-tuned LLM improves the precision of identifying privacy information in IoV traffic from 71.2% to 88.0%, and the recall rate from 85.7% to 86.3%. Among the 88 detected privacy leakage items, 32 items (accounting for 36.4%) are not stated in the devices privacy policies. More seriously, some IoV devices transmit captured video thumbnails and audio-to-text transcription results in unencrypted plaintext, causing severe risks of user personal privacy leakage.

To address the performance bottlenecks in IoT intrusion detection caused by resource-constrained devices, high-dimensional redundant traffic features, and imbalanced attack sample distributions, we propose an IoT intrusion detection method that integrates adaptive feature optimization and boundary-balanced sampling. First, to mitigate the increased computational overhead and classification interference caused by high-dimensional redundant features in IoT traffic data, we design an adaptive cross-validation recursive feature selection algorithm based on Light Gradient Boosting Machine (LightGBM), termed LightGBM-based Adaptive Cross-Validation Recursive Feature Selection (LARFS). This algorithm combines feature importance evaluation with a cross-validation recursive feature elimination mechanism and adaptively searches for the optimal feature subset, thereby effectively reducing feature dimensionality while maintaining detection performance. Second, to overcome the class imbalance in IoT traffic and the boundary noise introduced by conventional oversampling methods, we propose a data balancing method termed Weighted Borderline-SMOTE and Tomek-ENN-based Data Balancing (WBSTE), which combines weighted Borderline-Synthetic Minority Over-sampling Technique (weighted Borderline-SMOTE) with Tomek links and Edited Nearest Neighbors (ENN). By enhancing boundary samples, removing class overlap, and cleansing noise, WBSTE effectively improves the model’s ability to recognize minority attack samples. Experimental results on the N-BaIoT dataset demonstrate that the proposed method achieves high detection accuracy and low computational overhead, making it well-suited for resource-constrained IoT environments.

Most vulnerability detection methods for smart contracts rely on a single modality or large models. The former suffers from insufficient extraction of key features, resulting in low detection accuracy, while the latter has a large parameter scale, which is not conducive to deployment in resourceconstrained environments. To address these issues, this paper proposes dualmodal conditional feature gating, a lightweight dynamic feature fusion method for smart contract vulnerability detection, which adaptively generates fusion weights via a lightweight gating network and constructs a lightweight deploymentoriented detection framework. Experiments on 17,554 reentrancy samples from SmartBugs Wild achieve an accuracy of 76.87%, an F1score of 0.788, and an AUC of 0.870, outperforming static fusion, concatenation fusion, and singlemodal baselines. Compared with reproduced MVul-L and CMRL, the proposed method maintains comparable performance while reducing the parameter count by about 45% and lowering inference latency.

With the in-depth advancement of the “Internet+Government Services” initiative, e-government systems have an increasingly urgent need for secure authentication and trusted interaction. Traditional signature schemes based on cryptographic smart tokens suffer from prominent physical-media dependency, high operation-management costs and single-point-of-failure risks in scenarios such as distributed government collaboration and cross-departmental data sharing. This paper applies collaborative signature technology to e-government services and builds a novel signature system by integrating distributed key generation and threshold cryptography mechanisms. It focuses on the practical deployment of SM2 elliptic-curve public-key-algorithm-based collaborative signatures within e-government scenarios. On the basis of analyzing the limitations of conventional authentication and digital-signature modes, this paper systematically elaborates the theoretical foundation, technical architecture, implementation workflow and typical service scenarios of SM2 collaborative signatures for government businesses. The results demonstrate that the technology mitigates reliance on physical cryptographic media and improves key-management efficiency. It balances security and usability requirements for government services and offers new technical support for trusted interactions in digital government.

In recent years, with the rapid iterative development of information technology, tactics of telecom and cyber fraud have continued to evolve and upgrade. Fraudsters increasingly utilize generative artificial intelligence (AI) technologies such as deepfakes to fabricate or tamper with specific identity information with high fidelity, and carry out targeted telecom and cyber fraud on this basis. This issue has gradually become a prominent challenge in the prevention and governance of telecom and cyber fraud. To address these risks and problems, this research innovatively constructs a comprehensive multimodal deepfake detection system, which can identify AI-generated images, AI voice replacement, and AI face swapping. Meanwhile, an end-side lightweight multimodal detection SDK is also developed. The proposed system can be widely applied to a variety of business scenarios, including multimedia messages covering text, images, audio, video and virtual scenarios, 5G new calls, video conferences, and trusted identity authentication. It aims to provide flexible, efficient and accurate AI authenticity verification, early warning and authentication framework and capability for all industries.