广告
加载中

蚂蚁开源两款安全模型:50毫秒内截停AI闯祸

胡镤心 2026-07-13 16:15
胡镤心 2026/07/13 16:15

邦小白快读

EN
全文速览

本文核心信息是蚂蚁AI安全实验室7月13日开源两款AI安全护栏模型,回应AI能力升级后带来的新型安全挑战,核心干货如下:

1. 两款模型瞄准当前AI领域最突出的两类安全问题:SingGuard-NSFA服务自主执行任务的智能体,在工具调用、代码执行前完成安全检测;SingGuard服务多模态交互大模型,可识别跨模态隐藏的恶意攻击。

2. 两款模型都具备落地实操优势:SingGuard-NSFA最快50毫秒完成风险判定,支持轻量化部署,新增风险类别只需训练轻量模块,适配攻击手法快速变化的特点;SingGuard检测准确率超过Llama Guard 3、谷歌ShieldGemma等全球主流同类安全模型,支持不重新训练就更新安全规则,适配各类AI应用的安全需求。

当前AI已成为品牌数字化升级、产品创新的核心方向,本文带来AI布局层面的核心干货如下:

1. 政策与风险层面:国家已经明确将安全可控作为智能体发展的底线要求,品牌布局AI应用必须把安全放在合规前提,当前提示词注入、权限滥用、数据泄露等已经成为真实可发生的安全威胁,必须提前做好防护。

2. 产品研发层面:品牌布局智能体、多模态大模型应用时,可以直接利用蚂蚁开源的两款安全模型搭建安全防护体系,两款模型覆盖绝大多数风险场景,支持多语言,性能领先且适配轻量化部署,能降低品牌自身的AI安全研发成本,满足合规要求的同时保障AI应用稳定运行。

当前越来越多卖家开始布局AI应用提升运营效率,本文带来的相关干货如下:

1. 合规与风险提示:国家主管部门已经出台智能体发展相关规范,明确要求安全可控,卖家落地AI工具时必须满足合规要求,避免触碰监管红线;当前AI应用面临提示词注入、权限泄露、恶意攻击等真实风险,一旦出现问题会给店铺运营带来严重损失,必须提前防护。

2. 机会与可利用资源:蚂蚁开源的两款安全模型免费开放,覆盖智能体工具调用、多模态内容交互两大卖家常用AI场景,模型适配高并发线上场景,支持轻量化部署,卖家可以低成本将防护能力集成到自身使用的AI工具中,不用投入高额研发成本就能满足安全合规要求,降低AI应用的风险。

当前工厂推进数字化、智能化转型已经广泛应用AI技术,本文带来的干货启示如下:

1. 需求层面:工厂智能化转型过程中,AI越来越多地承担自主任务调度、多模态数据处理等工作,同样面临提示词注入、数据泄露、权限滥用等安全风险,随着监管要求明确,工厂落地AI应用必须将安全可控放在核心位置。

2. 转型启示:工厂推进AI落地不需要从零搭建安全防护体系,可以利用蚂蚁开源的两款成熟安全模型,两款模型适配不同规模的部署需求,支持轻量化部署,中小型工厂也能低成本应用,可覆盖工厂智能体调度、多模态质检等各类AI场景的安全需求,帮助工厂在推进数字化的过程中满足合规要求,规避安全风险。

作为为各行业提供AI相关服务的服务商,本文带来行业趋势、客户痛点与解决方案的核心干货如下:

1. 行业趋势与客户痛点:当前AI从能聊向能干升级,智能体自主执行任务、多模态交互已经成为主流方向,传统的关键词安全防护已经无法适配新型攻击方式,客户普遍存在AI安全防护的需求,同时攻击手法变化快,要求防护能力能快速迭代。

2. 成熟解决方案参考:蚂蚁开源的两款模型分别覆盖两大核心场景,SingGuard-NSFA从请求拦截和响应兜底搭建防护体系,支持快速迭代规则,适配高并发场景;SingGuard可识别跨模态隐藏攻击,检测性能领先全球主流产品,服务商可以将两款模型集成到自身的AI解决方案中,快速满足客户的AI安全合规需求,提升自身解决方案的竞争力。

面向AI开发者的平台商可从本文获得以下干货:

1. 当前平台入驻开发者的核心需求:越来越多开发者布局智能体、多模态大模型项目,普遍缺乏成熟的AI安全防护能力,同时监管明确要求AI应用安全可控,平台需要帮助开发者满足合规要求,规避整体生态的安全风险。

2. 平台生态建设方向:平台可以向开发者推广蚂蚁开源的两款安全模型,降低开发者的安全研发门槛,帮助开发者快速满足合规要求;平台自身也需要跟进AI安全监管风向,将安全能力作为AI开发者服务的基础配置,明确安全底线要求,完善平台AI生态的合规管理,避免行业风险传导到平台,同时吸引更多合规开发者入驻,丰富平台生态。

本文为AI产业、AI安全领域研究提供了以下新动向与研究参考:

1. 产业新动向:当前AI技术演进速度快于安全规则制定,AI安全已经成为制约行业发展的核心问题,蚂蚁作为国内头部科技企业,已经完成了AI安全领域的系统布局,先后开源多款安全防护模型,相关技术已经在自身支付等业务场景落地打磨,技术成熟度较高。

2. 研究方向参考:当前AI领域出现了新型攻击方式,攻击者转而攻击AI认知,带来了新的研究课题;政策层面已经出台智能体发展规范,明确安全底线,可研究政策落地对产业的影响;另外本文提出AI安全未来需要全行业共建公共安全基础设施,开源开放是AI安全基础设施的重要建设模式,这一方向也值得深入研究。

返回默认

声明:快读内容全程由AI生成,请注意甄别信息。如您发现问题,请发送邮件至 run@ebrun.com 。

我是 品牌商 卖家 工厂 服务商 平台商 研究者 帮我再读一遍。

Quick Summary

This article shares the core announcement that Ant Group's AI Security Lab open-sourced two AI safety guardrail models on July 13, developed to address emerging security challenges brought by advanced AI capabilities:

1. The two models target the two most pressing safety issues in the current AI sector. SingGuard-NSFA is designed for autonomous task-performing AI agents, conducting safety checks before they call tools or execute code. SingGuard caters to multimodal large language models, detecting hidden malicious attacks that span different modalities.

2. Both models offer practical advantages for real-world deployment: SingGuard-NSFA can complete risk judgment in as little as 50 milliseconds, supports lightweight deployment, and only requires lightweight fine-tuning for new risk categories, making it adaptable to rapidly evolving attack methods. SingGuard outperforms leading global safety models including Llama Guard 3 and Google's ShieldGemma in detection accuracy, and supports updating safety rules without retraining, adapting to the safety requirements of a wide range of AI applications.

Against the backdrop that AI has become a core driver of brands' digital transformation and product innovation, this article shares key takeaways for brands' AI strategy:

1. Policy and risk perspective: Chinese regulators have explicitly made safety and controllability the bottom line for AI agent development. Brands must prioritize safety to meet compliance requirements when rolling out AI applications. Today, threats such as prompt injection, privilege abuse, and data leakage are already tangible risks, and proactive protection is a necessity.

2. Product R&D perspective: When developing AI agent or multimodal large model applications, brands can build their safety protection systems directly using the two open-source models from Ant. The models cover most common risk scenarios, support multiple languages, deliver industry-leading performance, and support lightweight deployment. They can reduce brands' in-house AI security R&D costs, help meet compliance requirements, and ensure stable operation of AI applications.

As a growing number of sellers adopt AI to improve operational efficiency, this article shares key insights for sellers:

1. Compliance and risk reminder: Chinese regulators have released norms for AI agent development that explicitly require safety and controllability. Sellers must meet compliance requirements when deploying AI tools to avoid violating regulatory rules. AI applications currently face tangible risks including prompt injection, privilege leakage, and malicious attacks; security incidents can cause severe losses to store operations, so proactive protection is essential.

2. Accessible resources: The two AI safety models open-sourced by Ant are free to use, covering the two most common AI scenarios for sellers: agent tool calling and multimodal content interaction. The models are adapted for high-concurrency online scenarios and support lightweight deployment. Sellers can integrate protection capabilities into their AI tools at low cost, meeting safety and compliance requirements without heavy R&D investment, and reducing risks associated with AI adoption.

As factories widely adopt AI to advance digital and intelligent transformation, this article shares key takeaways:

1. Risk perspective: During the intelligent transformation process, AI is increasingly used for autonomous task scheduling and multimodal data processing, which also exposes factories to risks including prompt injection, data leakage, and privilege abuse. With clear regulatory requirements now in place, factories must prioritize safety and controllability when deploying AI applications.

2. Transformation insights: Factories do not need to build AI safety protection systems from scratch. They can leverage the two mature open-source safety models from Ant, which adapt to deployment needs of all scales and support lightweight deployment. Even small and medium-sized factories can adopt these models at low cost. The models cover safety needs for all common AI use cases in factories including intelligent agent scheduling and multimodal quality inspection, helping factories meet compliance requirements and mitigate security risks as they advance digital transformation.

As service providers that deliver AI-related solutions to various industries, you can find insights on industry trends, customer pain points, and solutions in this article:

1. Industry trends and customer pain points: AI is shifting from conversational capability to task execution capability, with autonomous AI agents and multimodal interaction becoming mainstream. Traditional keyword-based protection can no longer defend against new types of attacks. Clients universally demand AI safety protection that can iterate quickly to keep up with rapidly evolving attack methods.

2. Proven solution reference: Ant's two open-source models cover the two core AI safety scenarios respectively. SingGuard-NSFA builds a protection system through request interception and response fallback, supports rapid rule iteration, and adapts to high-concurrency scenarios. SingGuard can detect hidden cross-modal attacks and outperforms leading global products in detection performance. Service providers can integrate these two models into their own AI solutions to quickly meet clients' AI safety and compliance needs, and improve the competitiveness of their offerings.

AI developer platforms can draw the following key insights from this article:

1. Core needs of platform developers: A growing number of developers are working on AI agent and multimodal large model projects, but most lack mature AI safety protection capabilities. Meanwhile, regulators explicitly require AI applications to be safe and controllable, so platforms need to help developers meet compliance requirements to mitigate security risks for the entire ecosystem.

2. Directions for platform ecosystem development: Platforms can promote Ant's two open-source safety models to developers to lower the barrier for security R&D and help developers meet compliance requirements quickly. Platforms should also keep up with AI safety regulatory trends, integrate safety capabilities as a baseline feature for AI developer services, clarify bottom-line safety requirements, and improve compliance management for the platform's AI ecosystem. This will prevent industry risks from spilling over to the platform, attract more compliant developers, and enrich the platform ecosystem.

This article provides new industry trends and research references for researchers in AI industry and AI security:

1. New industry trends: AI technology is evolving faster than safety rulemaking, and AI security has become a core bottleneck restricting industry development. As a leading Chinese technology company, Ant has completed systematic布局 in the AI security field and has open-sourced multiple safety protection models. Its relevant technologies have been validated in internal production scenarios such as payment services, demonstrating high technical maturity.

2. Research direction references: New attack methods targeting AI cognition have emerged in the AI field, creating new research topics. Policymakers have released development norms for AI agents that clarify safety bottom lines, creating opportunities to study the impact of policy implementation on the industry. In addition, this article proposes that the future of AI security requires the whole industry to jointly build public safety infrastructure, and open-source is a key construction model for AI security infrastructure, a direction that merits in-depth research.

Disclaimer: The "Quick Summary" content is entirely generated by AI. Please exercise discretion when interpreting the information. For issues or corrections, please email run@ebrun.com .

I am a Brand Seller Factory Service Provider Marketplace Seller Researcher Read it again.

【亿邦原创】7月13日,蚂蚁AI安全实验室宣布开源智能体安全护栏SingGuard-NSFA,并同步披露多模态安全护栏SingGuard的详细信息。两款模型分别面向“自主执行的智能体”和“多模态交互的大模型”两大前沿场景,是蚂蚁在AI安全领域的系统布局。

过去一年,AI行业经历了一场从能聊到能干的跨越,智能体开始自主调用工具、运行代码、规划任务,模型也开始理解图像、文本、音频等多模态信息。

能力边界的拓展带来效率,也带来了新的安全挑战,提示词注入攻击从实验室概念变成了真实威胁,攻击者不再攻击代码,而是攻击AI的认知。权限滥用、数据泄露、恶意代码执行等事件接连发生,开源智能体OpenClaw带来更复杂的风险链条,攻击者可以通过22个以上频道入口进行提示词注入,可在60秒内完成从消息发送到shell再到root权限的跨越。

行业对风险的感知也在加速,2026年5月,国家网信办、国家发展改革委、工业和信息化部联合印发《智能体规范应用与创新发展实施意见》,明确将“安全可控”作为智能体发展的底线要求。政策与标准在追赶技术,但AI的演进速度远快于规则制定——安全能力需要同步进化。

蚂蚁AI安全实验室此次开源的两款模型,正是对这种需求的具体回应。

SingGuard-NSFA瞄准的是智能体动手那一刻的安全。当智能体准备调用工具、执行代码或编排多步任务时,SingGuard-NSFA在动作执行前完成实时安全检测,从请求拦截和响应兜底两端构建行为安全防护体系。它基于CIA(保密性、完整性、可用性)原则,结合OWASP等国际安全指南,将智能体风险细分为7大类、28个中类和185个具体场景,并建立覆盖133种语言、近10万条样本的评测体系。

在技术实现上,SingGuard-NSFA提供两种工作模式。

一种模式逐条生成详细的风险分析报告,方便事后审查;另一种模式在50毫秒左右完成单次风险判定,适合线上高并发场景下的实时拦截。它提供0.8B、2B、4B、9B四种模型规模,其中0.8B模型即可达到8B模型的性能水平,轻量化部署成为可能。新增风险类别时仅需训练轻量模块,无需重新训练整个模型——这种设计思路本身就是在回应一个现实:AI的攻击手法每天都在变,安全能力必须跟得上变化的速度。

SingGuard则面向多模态交互场景下的内容安全。

今年6月,Anthropic发布Claude Fable 5,数日内即被研究者用Unicode字符和西里尔字母替换敏感词绕过了安全护栏——模型能还原原意,分类器却视其为陌生拼写。这意味着模型越能读懂变形文字、图片和跨模态信息,传统关键词识别式的护栏就越力不从心。

SingGuard正是为此设计,它面向文本、图片及跨模态内容建立统一安全判断框架,可识别攻击者将恶意动作隐藏在文字、图片等不同模态中的复杂攻击。同时支持运行时动态加载自然语言安全规则,无需重新训练模型即可完成规则更新。

在推理机制上,SingGuard采用“快慢结合”模式:先快速完成初步判断,只有面对复杂场景时才启动进一步推理,在保证效率的同时提升检测准确率。

在覆盖文本查询、文本回复、图像、多模态和多语言的6大类评测中,SingGuard在35个数据集及评测切分上的平均F1均为最高,对比对象涵盖Llama Guard 3、谷歌ShieldGemma、GPT-5.1、Gemini3-Pro等主流护栏,SingGuard均实现全面领先。

蚂蚁在AI安全领域的布局并非一日之功。今年4月,蚂蚁AI安全实验室联合清华大学开源了智能体安全防御插件ClawAegis,为自主智能体提供覆盖全生命周期的安全防护能力。

此次SingGuard-NSFA与SingGuard相继开源,是蚂蚁在AI安全领域的又一次系统性输出。

这些技术积累建立在蚂蚁二十余年的安全实践上,面向支付安全、数据安全、隐私保护、风险治理等真实业务的长期打磨,相关能力已应用于蚂蚁阿福、AI版支付宝阿宝、支付宝AI付等业务场景。

同时,蚂蚁持续参与IIFAA《终端智能体可信互联技术规范》制定,牵头ITU国际标准《终端智能体可信互联技术规范》立项,并发布智能体安全可信互连协议ASL。

AI时代的安全体系需要将防御嵌入AI的每一次推理、每一次工具调用、每一次决策之中。当越来越多的开发者将SingGuard和SingGuard-NSFA集成到自己的AI应用中,这些模型才能在真实攻击中不断进化。AI安全的未来,不取决于某一家公司的技术有多强,而取决于整个行业能不能建立起一套共同的安全基础设施。

文章来源:亿邦动力

广告
微信
朋友圈

FAQ回顾

SingGuard-NSFA安全模型有哪些核心优势?

SingGuard-NSFA是蚂蚁开源的智能体安全护栏,可在智能体执行动作前完成实时安全检测,单次风险判定仅需50毫秒左右适配高并发场景,提供4种模型规模,0.8B版本即可达到8B模型的性能水平,新增风险仅需训练轻量模块。

多模态大模型安全护栏SingGuard有什么特点?

SingGuard是蚂蚁开源的多模态安全护栏,建立跨模态统一安全判断框架,可识别隐藏在不同模态中的复杂攻击,支持动态加载安全规则,在6大类35个数据集评测中平均F1值最高,全面优于Llama Guard 3等多款主流安全护栏。

我国对智能体行业发展有什么底线要求?

2026年5月,国家网信办、发改委、工信部联合印发《智能体规范应用与创新发展实施意见》,明确将安全可控作为智能体发展的底线要求,行业安全能力需匹配AI技术的快速演进节奏。

这么好看,分享一下?

朋友圈 分享

APP内打开

赞 +1
+1
微信好友 朋友圈 新浪微博 QQ空间
关闭
收藏成功
发送
/140 0