内容来源：https://aibusiness.com/foundation-models/microsoft-cheaper-cyber-model-to-rivals

内容总结：

微软推出低成本AI网络安全模型，剑指企业级生成式AI应用场景

微软近日正式发布新一代网络安全模型MAI-Cyber-1-Flash，同时推出名为“Project Perception”的智能体安全平台，旨在以更低成本满足企业对生成式AI服务的旺盛需求。此举被业界视为微软在人工智能战略调整中的关键落子，此前其Copilot系统在企业端表现平平。

据微软介绍，MAI-Cyber-1-Flash模型综合调用OpenAI旗下三款AI模型（GPT-5.4、GPT-5.4及GPT-5.3-Codex），可智能路由安全漏洞识别请求，其运行成本仅为竞品系统的一半。该模型嵌入于今年5月发布的多智能体漏洞识别与修复平台MDASH中，而Project Perception则为MDASH提供红、蓝、绿三色安全团队智能体，协同完成漏洞监控与修复工作。

在性能方面，微软声称MAI-Cyber-1-Flash在广泛使用的行业基准测试中，表现超越了Anthropic的Mythos、OpenAI的GPT-5.6 Sol以及谷歌的Gemini 3.5 Flash Cyber等主流模型。值得注意的是，此前Anthropic的Mythos模型因被特朗普政府认定为国家安全威胁而撤出市场，OpenAI的GPT-5.6也通过“Project Daybreak”计划仅向受许可用户开放。微软则强调，其新模型面向所有企业开放，且通过智能路由技术大幅降低token消耗成本。

分析人士指出，微软此举充分借助了自身在企业软件安全领域的长期积累。Futurum集团分析师David Nicholson认为，微软模型本质上是“按需路由”架构，针对不同安全任务选择最低成本的模型处理，而非像竞品那样提供“超豪华跑车”式的全功能模型。Forrester分析师Allie Mellen则补充道，企业构建智能体系统真正的难点不在于模型本身，而在于围绕模型的整合框架。微软通过垂直整合优势，将自身数据与专业经验注入模型，使其更适配微软生态内的安全场景，为用户节省了架构设计与资源调配的时间成本。

随着企业级生成式AI服务降本需求的急剧攀升，微软正以“低成本+全栈整合”策略重仓AI网络安全赛道，而其能否凭借此轮布局扭转此前Copilot在企业端的平淡表现，仍有待市场检验。

中文翻译：

由谷歌云赞助
选择首个生成式AI应用场景
要开始使用生成式AI，首先应聚焦于那些能改善人类与信息交互体验的领域。

随着该公司加速推进AI战略，这家科技巨头正在竞争激烈的AI网络安全领域积极布局。

由于企业对低成本生成式AI服务的需求急剧攀升，微软推出了一款网络安全模型和平台，声称其成本仅为竞品系统的一半。

这项网络安全举措出台之际，微软正试图重振其AI策略——此前其Copilot AI系统在企业市场反响平淡。紧随其后的是一系列近期发布的智能体AI产品，以及在全球范围内建设和运营AI数据中心的大规模计划。

微软还宣称，其于周一发布的MAI-Cyber-1-Flash模型，在一项广泛使用的基准测试中表现优于Anthropic的Mythos、OpenAI的GPT-5.6 Sol以及谷歌的Gemini 3.5 Flash Cyber。

这款网络安全模型将安全漏洞识别请求路由至OpenAI的三款AI模型，并嵌入微软于5月发布的MDASH多智能体漏洞识别与修复平台。

与MAI-Cyber-1-Flash一同推出的还有Project Perception，这是一个智能体安全平台，可为MDASH中的各种安全流程提供智能体团队，用于监控和修补安全漏洞。微软表示，除了最初的软件漏洞识别应用外，近期还将在Perception中集成MAI-Cyber-1-Flash，以支持更多安全流程。

微软此次网络安全行动，正值Anthropic的Mythos模型引发争议之际。该模型曾被特朗普政府视为国家安全威胁并强制其退出市场，随后被纳入Anthropic于4月发布的Project Glasswing项目，仅限特定组织使用。OpenAI的主要网络安全模型GPT-5.6也经历了类似轨迹。尽管政府未对其施压，但OpenAI在5月将其纳入Project Daybreak项目，同样仅限授权用户使用。

“显然，这还是个‘有待观察’的领域。但我认为关键在于，它面向所有人开放，而不像Mythos和GPT-5.6那样受限，”Futurum Group分析师戴维·尼科尔森表示，“而且它的成本要低得多，因为它能智能地将令牌消耗路由至合适的模型。”

MAI-Cyber-1-Flash目前基于OpenAI的GPT-5.4、GPT-5.4和GPT-5.3-Codex构建——这些模型除提供推理、编码和智能体工作流等生成式AI功能外，还具备网络安全能力。

尼科尔森表示，本质上MAI-Cyber-1-Flash本身就是一个路由模型，能为特定安全任务选择最低成本的模式，而非像Mythos和GPT-5.6这类功能强大但价格高昂的专用网络安全模型——他将其比作高端、超贵的跑车。

“他们传达的信息是：‘我们会在模型间做套利，为合适的任务挑选合适的模型，这样你就不用花冤枉钱’，”他说。

尼科尔森指出，微软还在发挥自身优势：长期专注于企业软件安全。由于微软平台在企业IT中无处不在，其系统构成了巨大的攻击面，多年来一直是网络攻击的重点目标；与此同时，微软也构建了强大的安全机制来保护其软件。

Forrester分析师阿莉·梅伦表示，Project Perception通过红队、蓝队和绿队智能体的协作，为这些模型提供了协调不同团队智能体的框架。

“它负责编排这些智能体……选择最佳平衡质量与成本的模型，并为系统提供恰当的上下文以获得最优结果，”她说，“构建智能体系统真正的难点不在于模型本身，而在于围绕它的框架。微软将这套完整系统作为产品发布，可以节省用户的时间、资源和架构设计成本。”

梅伦和尼科尔森均表示，微软方法的另一个优势在于其垂直整合能力。

“微软推出自有模型，因此对整个技术栈拥有更强掌控力，”梅伦说，“基于自身数据和专业知识的模型，能确保其最擅长推理微软数据，并与其产品高度契合。”

英文来源：

Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
As it ramps up its AI strategy, the tech giant is making a play in the hotly contested AI cybersecurity arena.
With enterprise demand for lower-cost generative AI services spiking sharply, Microsoft introduced a cybersecurity model and platform it said costs half that of rival systems.
The cybersecurity initiative comes as Microsoft aims to resuscitate its AI approach after its Copilot AI system was met with lukewarm success by enterprises. It follows a series of recent agentic AI releases and a large-scale undertaking to build and operate AI data centers worldwide.
Microsoft also claimed that its MAI-Cyber-1-Flash model, released on Monday, outperformed Anthropic's Mythos, OpenAI's GPT-5.6 Sol and Google's Gemini 3.5 Flash Cyber on a widely used benchmark.
The cybersecurity model, which routes security vulnerability identification requests to three AI models from OpenAI, is embedded in the vendor's MDASH multi-agent vulnerability identification and remediation platform, which it released in May.
Alongside MAI-Cyber-1-Flash, Microsoft launched Project Perception, an agentic security platform that furnishes teams of agents for various security workflows in MDASH to monitor and patch security vulnerabilities. Microsoft said it will soon add MAI-Cyber-1-Flash to Perception for many more security workflows in addition to its initial software vulnerability identification application.
Microsoft's cybersecurity move comes in the wake of furor over Anthropic's Mythos model, which the Trump administration initially deemed a national security threat and forced it off the market. That model was subsumed under Anthropic's Project Glasswing, launched in April, which restricted Mythos to select organizations. OpenAI's main cyber model, GPT-5.6, followed a similar trajectory. Although the government did not clamp down on it, in May, OpenAI made it part of Project Daybreak, which is similarly restricted to approved users.
"Obviously, this is 'remains to be seen' territory. But I think the key here is that it's available to everybody, as opposed to Mythos [and GPT 5.6], which is not," said David Nicholson, an analyst at The Futurum Group. "And it's going to cost you a lot less because it's intelligently routing the consumption of tokens to appropriate models."
MAI-Cyber-1-Flash is currently built on OpenAI's GPT-5.4, GPT-5.4 and GPT-5.3-Codex -- all of which provide cybersecurity capabilities in addition to generative AI features such as reasoning, coding and agentic workflows.
In essence, Nicholson said, MAI-Cyber-1-Flash is itself a routing model that selects the lowest-cost mode for specific security tasks, as opposed to powerful, cybersecurity-dedicated models such as Mythos and GPT-5.6, which he likened to a high-end, ultra-expensive sports car.
"They're saying, 'We will do this arbitrage between models to pick the right model for the right job, so that you're not paying top dollar,'" he said.
Microsoft is also playing to one of its strengths: a longtime focus on enterprise software security, Nicholson said. Because of its pervasive presence in enterprise IT, Microsoft platforms have created a huge attack surface and for years have been a favored target for cyberattacks; at the same time, Microsoft has built formidable security mechanisms to try to protect its software.
Meanwhile, Project Perception, with its red, blue and green team agents, provides a harness around the models that coordinates agents from different teams, said Allie Mellen, a Forrester analyst.
"It is handling the orchestration of these agents ..., the choice of model that best balances quality and cost and is able to provide the right context to the system to get the best outcome," she said. "The truly difficult part of building an agentic system is not the model itself; it's the harness around it. Microsoft is releasing that comprehensive system as a product, which can save users time, resources and architecting."
Another advantage of Microsoft's approach is that it benefits from the Microsoft's vertical integration, Mellen and Nicholson said.
"Microsoft is launching its own model, so it has more control over the entire stack," Mellen said. "Its own model, based on its own data and expertise, ensures the model is best suited to reason over Microsoft data and best aligns to its products."