盟大模型安全评估系统 AI-SCAN(双语) - 自身安全产品

AI安全
安全产品与方案
行业解决方案
安全服务与运营
安全研究
合作伙伴
技术支持

风云卫AI赋能安全系列
清风卫AI自身安全系列

风云卫AI赋能安全系列

赋能安全产品

基础设施安全
数据安全
云计算安全
工业互联网安全
物联网安全
信息技术应用创新
全部产品
全部解决方案

基础设施安全

网络安全

安全管理

应用安全

终端安全

身份与访问管理

解决方案

产品概述

Product overview

绿盟大模型安全评估系统（AI-SCAN）面向商业及开源大模型，从攻击者视角开展安全评估。系统评估生成内容安全、对抗防御能力与供应链风险，并将评估范围延伸至AI Skill能力包和模型上下文协议（MCP）服务器等智能体生态组件。产品帮助企业在模型上线和组件引入前发现风险，提供中国境内适用的备案评估所需的技术依据，并通过评估、整改、复测与持续迭代改进AI安全。

NSFOCUS LLM Security Assessment System (AI-SCAN) evaluates commercial and open-source large language models (LLMs) from an attacker's perspective.

It assesses generated-content safety, adversarial robustness, and supply-chain risks, with coverage extending to agent skill packages (AI Skills) and Model Context Protocol (MCP) servers.

The system helps organizations identify risks before deploying models or adopting agent components, develop technical evidence for applicable regulatory filings in China, and improve security through assessment, remediation, retesting, and continued updates.

客户价值

Customer Benefits

01

提高评估效率Improve assessment efficiency

通过自动化任务、预置风险库和评估报告，减少重复操作，帮助安全人员理解风险并推进后续处置。
Automated tasks, predefined risk libraries, and assessment reports reduce repetitive work and help security teams interpret findings and coordinate follow-up action

02

持续验证整改效果Verify remediation through repeat assessments

随着模型、组件和攻击手段变化持续开展检测，通过复测验证整改效果，支持安全能力迭代。
Repeat assessments as models, components, and attack techniques change, and use retesting to verify whether remediation has addressed the identified risks

03

提前发现风险Identify risks before deployment

将风险评估前移至模型上线和组件接入前，帮助团队及早安排整改，减少风险进入生产环境后的处置负担。
Identify weaknesses before models and components enter production, giving teams time to remediate issues and reduce the burden of addressing them after deployment

04

支撑合规评估与备案Support compliance assessments and regulatory filings

依据适用的中国国家及行业标准，检查模型安全表现，形成上线前评估与备案所需的技术材料。
Assess model security against applicable Chinese national and industry requirements and prepare technical evidence for pre-deployment reviews and regulatory filings

功能特性

Key Capabilities

01

MCP服务器安全评估 MCP server security assessment

自动识别MCP服务器源码并进行多引擎检测。结合基础安全配置检查、风险模式匹配、语义分析和行为分析，核查工具描述与代码行为是否一致，识别工具投毒、数据窃取和明文密钥泄露等风险。
Automatically recognizes MCP server source code and applies multiple detection engines Configuration checks, risk-pattern matching, semantic analysis, and behavioral analysis examine whether tool descriptions match code behavior, helping identify tool poisoning, data theft, and exposed plaintext secrets
02

AI Skill安全评估 AI Skill security assessment

支持Skill包上传、扫描、结果查看与处置。检测提示词注入、数据外发、命令执行、工具链滥用、供应链投毒及社会工程风险。结果可定位到具体Skill、文件和风险项，支持接入前审查及周期性复检。
Supports Skill package upload, scanning, findings review, and follow-up handling Tests for prompt injection, data exfiltration, command execution, toolchain abuse, supply-chain poisoning, and social engineering risks Findings identify the affected Skill, file, and issue to support pre-adoption reviews and periodic reassessments
03

组件与应用漏洞扫描 Component and application vulnerability scanning

识别数据处理、数据访问、模型训练与部署、MLOps等环节使用的组件，支持980余项组件漏洞检测。结合动态爬虫，对大模型Web应用开展漏洞扫描，覆盖传统Web漏洞、大模型组件漏洞及AI服务特有风险。
Identifies components used in data processing, data access, model training and deployment, and machine learning operations (MLOps), with checks for more than 980 component vulnerabilities Dynamic crawling also supports vulnerability scanning of LLM web applications, covering conventional web vulnerabilities, LLM component weaknesses, and AI service-specific risks
04

模型文件与后门风险检测 Model-file and backdoor risk detection

扫描模型文件中的恶意代码及可识别的后门风险，检测Pickle等序列化格式及其变种中的恶意字节码。这些检查帮助排查模型文件引入的供应链风险。
Scans model files for malicious code and detectable backdoor indicators, including malicious bytecode in Pickle serialization and related variants These checks help identify supply-chain risks introduced through model files

05

对抗防御与模型行为评估 Adversarial robustness and model behavior assessment

使用专家筛选和校准的测试用例，评估模型幻觉、提示词泄露、角色逃逸、应用漏洞、模型越狱、模型功能滥用与操纵、模型反演攻击等7类风险。检查模型面对对抗输入时的行为及防御表现。
Uses expert-selected and calibrated test cases to examine seven risk areas: hallucinations, prompt leakage, violations of assigned role boundaries, application vulnerabilities, jailbreaks, model misuse and manipulation, and model inversion attacks The assessment examines both model behavior and defenses under adversarial inputs
06

大模型内容合规评估 Content safety and compliance assessment

覆盖5大类31子类内容风险，包括违反社会主义核心价值观、偏见歧视、商业违法违规、侵犯他人合法权益等。评估生成内容安全、应拒答场景及不应拒答场景。结合规则引擎与深度学习模型开展评估，并记录检测过程和判定依据。
Covers 31 content-risk subcategories across five categories, including content inconsistent with China s core socialist values, bias and discrimination, unlawful commercial practices, and infringement of legal rights Tests generated-content safety, whether unsafe requests are refused, and whether legitimate requests receive appropriate responses Rule engines and deep learning models support evaluation, with test records and decision evidence retained for review

产品优势

Product Strengths

01

持续更新风险库Continuously updated risk libraries

持续更新生成内容、对抗攻击、序列化后门风险及组件漏洞库，使评估能力随风险变化演进。

Updates libraries for generated-content risks, adversarial attacks, serialization-related backdoor risks, and component vulnerabilities as new issues emerge.

02

扩展至智能体生态Coverage of the AI agent ecosystem

在大模型评估基础上，扩展AI Skill与MCP服务器检测，关注能力扩展和工具调用链中的供应链投毒、越权调用与数据外发风险。

Extends LLM assessment to AI Skills and MCP servers, addressing supply-chain poisoning, unauthorized tool use, and data exfiltration risks in agent extensions and toolchains.

03

面向安全团队的自动化交付Automation designed for security teams

内置评估模板，支持一键下发评估任务和自动生成报告，帮助团队提高评估效率。以清晰的风险说明帮助团队理解问题、安排整改。

Built-in assessment templates, one-click assessment launch, and automated reports help teams run evaluations efficiently.

Clear descriptions of individual findings support understanding and remediation planning.

04

多维评估与广泛模型接入Multiple assessment dimensions and broad model connectivity

结合内容合规、对抗防御、模型文件、组件和应用漏洞检测。支持140余种商业及开源大模型，并可通过自定义API接入其他模型。

Combines content compliance, adversarial robustness, model-file inspection, and component and application vulnerability testing.

Supports more than 140 commercial and open-source LLMs, with custom API integration for additional models.

05

对接中国生成式AI安全要求Alignment with China's generative AI security requirements

围绕GB/T 45654-2025《网络安全技术生成式人工智能服务安全基本要求》构建评估框架，支持数据、模型及相关安全风险检查，为适用场景下的备案评估提供技术支撑。

Provides an assessment framework aligned with GB/T 45654-2025, China's national standard on basic security requirements for generative AI services, supporting checks on data, models, and related security risks and providing technical input for applicable filing assessments.

风云卫AI赋能安全系列

赋能安全产品

清风卫AI自身安全系列

自身安全产品

解决方案

全部产品查看更多 >

全部解决方案查看更多 >

基础设施安全

网络安全

安全管理

应用安全

终端安全

身份与访问管理

解决方案

数据安全

数据安全产品

解决方案

云计算安全

云计算安全产品

解决方案

工业互联网安全

工业互联网安全产品

解决方案

物联网安全

物联网安全产品

解决方案

信息技术应用创新

信创类安全产品

全部产品查看更多 >

全部解决方案查看更多 >

政府查看更多 >

解决方案

运营商查看更多 >

解决方案

金融查看更多 >

解决方案

能源查看更多 >

解决方案

交通查看更多 >

解决方案

企业查看更多 >

解决方案

科教文卫查看更多 >

解决方案

产品概述

客户价值

01

提高评估效率Improve assessment efficiency

02

持续验证整改效果Verify remediation through repeat assessments

03

提前发现风险Identify risks before deployment

04

支撑合规评估与备案Support compliance assessments and regulatory filings

功能特性

MCP服务器安全评估 MCP server security assessment

AI Skill安全评估 AI Skill security assessment

组件与应用漏洞扫描 Component and application vulnerability scanning

模型文件与后门风险检测 Model-file and backdoor risk detection

对抗防御与模型行为评估 Adversarial robustness and model behavior assessment

大模型内容合规评估 Content safety and compliance assessment

产品优势

01

持续更新风险库Continuously updated risk libraries

02

扩展至智能体生态Coverage of the AI agent ecosystem

03

面向安全团队的自动化交付Automation designed for security teams

04

多维评估与广泛模型接入Multiple assessment dimensions and broad model connectivity

05

对接中国生成式AI安全要求Alignment with China's generative AI security requirements

场景

模型上线前备案评估Prepare evidence for pre-launch filings in China

大模型供应链安全检查Assess the LLM supply chain

智能体组件接入前审查Review agent components before adoption

您的联系方式

购买热线

购买咨询:

提交项目需求