Paper 1: A Survey of Architectural Security in Harnesses for Large Language Model Agents

针对大模型智能体通过工具调用改变真实环境时面临的跨部件安全风险，文章以连接模型与环境的运行支撑层Harness为研究对象，采用定向文献检索与引用追踪，梳理1974---2026年的223项资料。围绕运行职责、威胁描述、保障机制及评价与演化建立分析框架，将七类运行部件与攻击者能力、策略实施位置和证据类型交叉标注，比较核心机制的保护范围、可信前提及效用代价。综合分析表明，来源区分、任务授权、环境隔离与独立验证分别约束不同环节；跨部件防护需要保留来源信息，将授权绑定具体操作与资源，并维持检查与提交的状态一致性。组件组合验证、安全与效用联合评估以及运行架构演化治理是后续研究的主要方向。

This survey examined cross-component security risks that arise when large language model agents change real environments through tool use. It focused on the harness, the runtime layer connecting a model to its environment, and reviewed 223 sources from 1974 to 2026 through targeted literature searches and citation tracing. It developed an analytical framework covering runtime responsibilities, threat descriptions, security mechanisms, and evaluation and evolution. It organized harness responsibilities into seven components: context management, tool interfaces, execution environments, control loops, state and memory, permissions and human oversight, and verification and feedback. It cross-mapped these components to attacker capabilities, enforcement locations, and evidence types, and compared the protection scope, trust assumptions, and utility costs of key defenses. The synthesis identifies distinct roles for provenance separation, task authorization, environment isolation, and independent verification. Cross-component protection requires preserving provenance, binding authorization to specific operations and resources, and maintaining state consistency between checks and commits. Compositional verification, joint evaluation of security and utility, and governance of harness evolution are key directions for further research.

Paper In "信息网络安全 ISSN1671-1122"｜Click Here To Get The Full Paper

Paper 2: ITRAQ and PRM-based quantitative saliva proteomics in gastric cancer: biomarker discovery

Background: Gastric cancer (GC) is a major global health issue with high mortality rates, primarily because over 80% of patients are diagnosed at an advanced stage. Current diagnostic methods, such as endoscopy, are invasive, while conventional tumor markers lack sensitivity for early detection. Salivary proteomics presents a promising, non-invasive, and cost-effective strategy for biomarker discovery.

1. This study aimed to identify differentially expressed salivary proteins (DEPs) in GC patients and evaluate their potential as novel non-invasive biomarkers.

2. Methods: Salivary proteomes from two groups of GC patients (group 1: n = 12; group 2: n = 13) and healthy controls (n = 11) were analyzed using isobaric tags for relative and absolute quantitation (iTRAQ). DEPs were identified and subjected to functional annotation (GO, KEGG) and protein-protein interaction (PPI) network analysis. Candidate biomarkers were subsequently validated using parallel reaction monitoring (PRM).

3. Results: A total of 671 proteins were identified. We found 124 and 102 significant DEPs in GC groups 1 and 2, respectively, compared to controls. Fifty-six overlapping DEPs were detected between the two GC cohorts, comprising 24 upregulated and 32 downregulated proteins. Functional analysis and PRM validation highlighted four key proteins with consistent differential expression: S100A8, S100A9, CST4, and CST5.

4. Interestingly, while CST4 and CST5 were downregulated in the saliva of GC patients, they are reportedly upregulated in GC tissue and blood.

5. Conclusion: Our findings demonstrate that specific salivary proteins, particularly S100A8, S100A9, CST4, and CST5, hold significant potential as a panel of non-invasive biomarkers for gastric cancer detection. These results provide new insights into saliva-based diagnostics and underscore the importance of cross-comparing biomarker expression across different biofluids and tissues.

Paper In "Frontiers in Molecular Biosciences"｜Click Here To Get The Full Paper

Paper 3: The Performance of AI in China’s National Civil Service Exam (NCSE)

This study explores the potential of large language models, specifically ChatGPT-4, in the context of high-stakes exams by evaluating its performance on the Chinese National Civil Service Exam (NCSE). With the rapid advancements in artificial intelligence, understanding how well AI can perform on exams designed to test comprehensive cognitive skills has become increasingly relevant. Inspired by previous studies that assessed ChatGPT’s performance on the U.S. Bar Exam, we aimed to extend this evaluation to measure its “intelligence quotient” through the NCSE, a more diverse and demanding benchmark than traditional IQ tests.

The NCSE includes the Administrative Aptitude Test (AAT) and Argumentative Essay Writing (AEW), covering a range of cognitive domains such as logical reasoning, reading comprehension, quantitative analysis, and memory-based questions. This structure allows for a holistic assessment of ChatGPT-4’s capabilities, including language processing, logical reasoning, mathematical calculations, and data analysis. Results indicate that ChatGPT-4 demonstrates considerable strengths in natural language comprehension and structured writing but reveals notable limitations in visual recognition and logical reasoning, especially in tasks requiring abstract thought and multi-step problem-solving.

Using the Fenbi grading platform for evaluation, ChatGPT-4’s scores were compared against average human test-takers, providing a reliable benchmark of the model’s performance relative to human standards. The study shows that ChatGPT-4 exceeds the human average in certain areas, yet falls short of the highest human scores, underscoring the need for continued development in AI’s logical reasoning and response regulation capabilities.

Our findings suggest that with ongoing advancements, AI models like ChatGPT-4 could potentially serve as valuable tools in academic and professional assessments. The NCSE offers a robust framework for evaluating AI’s practical cognitive skills, marking an innovative step in redefining intelligence metrics for AI in complex, real-world scenarios. This research contributes to the growing body of knowledge on AI assessment, setting the foundation for future applications and improvements in AI-driven evaluation systems.

Click Here To Get The Full Paper In CN｜Click Here To Get The Full Paper In US