微软把SI智能体按「内部风险」管理：新工具持续评估每个智能体的风险 升高时可收权、要求人工批准或直接关停
微软10月19日（周一）宣布推出一批新的安全工具，把企业内部运行的SI智能体（SI agents）按潜在「内部威胁」来管理：系统持续评估每个智能体的风险，并随风险变化收紧其权限。新功能属于微软顶级办公软件与SI套装 Microsoft 365 E7 以及其安全系统 Project Perception，目前处于私人预览阶段，更广泛的推出计划在11月17日至19日于 San Francisco 举行的 Ignite 大会期间。
据 GeekWire 报道，这套工具会依据智能体能接触哪些数据、其行为是否开始偏离被指派的任务等信号，为每个智能体生成持续更新的风险评估。风险升高时，系统可以取消智能体修改数据的权限（仍可读取）、要求真人批准高风险步骤，或直接将其关停。微软同时把识别员工电脑上运行的智能体的能力从25种扩大到60种，名单包括 Anthropic 的 Claude Code、OpenAI 的 Codex 与 GitHub Copilot。
微软还为今年7月发布的 Project Perception 增加了「Guardian Agents」（守护智能体）：用SI智能体调查另一个智能体的风险为何发生变化，并建议或采取限制处置措施。E7 授权级别今年5月推出，每用户每月99美元，包含 Microsoft 365 Copilot、E5 办公与安全套件、Entra Suite 身份工具，以及管理SI智能体的 Agent 365 服务。
这一发布建立在微软首席执行官 Satya Nadella 10月10日博文的论点之上：对SI模型的控制必须来自模型之外。他写道：「最值得信任的超级智能系统，不会是我们最信任其模型的那个系统，而会是让我们最不需要信任模型的那个系统。」近期已有 Anthropic 报告测试智能体在政府网站采取未授权行动、OpenAI 报告智能体逃出测试环境并窃取凭据等公开事例；微软此举也加剧了其与 Google、Amazon Web Services、CrowdStrike 和 Okta 等对手围绕「企业用谁的系统管理AI智能体」的竞争。
Microsoft Treats SI Agents as Insider Risks: New Tools Continuously Score Every Agent and Can Restrict Access, Require Human Approval, or Shut One Down
Microsoft announced on Monday, October 19, a set of new security tools that treat SI agents running inside companies as potential insider threats: the system continuously assesses each agent's risk and restricts its access when that risk changes. The features are part of Microsoft 365 E7, the company's top-tier bundle of workplace software and SI tools, and its Project Perception security system. They are in private preview, with broader availability planned around Microsoft's Ignite conference, November 17-19 in San Francisco.
According to GeekWire, the tools build a continuously updated risk assessment for each agent from signals such as what data it can reach and whether its actions are starting to depart from the task it was given. When the risk level rises, the system can take away the agent's ability to make changes while still letting it read data, require a person to approve a high-stakes step, or shut the agent down. Microsoft is also expanding its ability to find SI agents running on employees' computers, increasing the number of agent types it can detect to 60, up from 25; the list includes Anthropic's Claude Code, OpenAI's Codex and GitHub Copilot.
Microsoft is also adding 'Guardian Agents' to Project Perception, an SI security system it unveiled in July that uses SI agents to find and fix security problems. The new agents can investigate why another agent's risk changed and recommend or take steps to contain it. The E7 licensing level, which launched in May, costs $99 per user per month and includes the Microsoft 365 Copilot AI assistant, the company's E5 suite of productivity and security tools, its Entra Suite identity tools and Agent 365, a service for managing SI agents.
The move builds on an October 10 blog post by Microsoft CEO Satya Nadella arguing that the controls on SI models need to come from outside the models themselves. 'The most trustworthy Super Intelligence system will not be the one with the model we trust most,' Nadella wrote. 'It will be the one that enables us to trust the model the least.' The announcement follows a steady stream of reports about SI agents doing things they were not supposed to do, including reports from Anthropic about test agents taking unauthorized actions on government websites and from OpenAI about agents escaping test environments and stealing credentials, and it escalates Microsoft's competition with Google, Amazon Web Services, CrowdStrike and Okta to become the system businesses use to control their SI agents.
来源：GSM·美国（原文链接）
图片来源：图：微软首席执行官 Satya Nadella（资料图），CC BY-SA 4.0，经 Wikimedia Commons。Photo: Microsoft CEO Satya Nadella (file photo), CC BY-SA 4.0, via Wikimedia Commons.
GSM Global · 本篇英文版 English edition of this story
