微软升级Project Perception 模型不可信

时间:2026-10-19 地点:美国华盛顿州雷德蒙德 人物:Satya Nadella(微软 CEO)、Charlie Bell(微软安全首席)、Vas Vasquez握手为Vasu Jakkal(微软安全营销 VP) 事件:微软宣布 Project Perception 重大升级,CEO 萨提亚·纳德拉公开表态"前沿 AI 模型无论来源,一律不可信"(frontier models, regardless of their origin, cannot be trusted),并把 AI 代理正式定性为"内鬼威胁"(insider risk);同步推出 Guardian Agents(守护代理)作为新一层监督机制。 影响: - 微软成为首家公开把"前沿模型不可信"写入产品哲学的云厂商,直接给企业级 AI 部署定下新的安全基线 - Guardian Agents 以代理监督代理,代表"agentic security"(代理化安全)进入主流,与 AgentOps、Red-Blue-Purple 演练并列 - Microsoft 365 E7 与 Defender 等安全 SKU 集成,带动一波企业合规与零信任升级需求 - 把 AI 风险类比为 insider threat,直接抬高企业 CISO 责任边界,可能影响董事会的 AI 采购审批流程 总结: 微软用 7 月首发的 Project Perception 作为底座,在 10-19 把"前沿模型不可信"从口号落到工具:Guardian Agents 持续评估代理行为、动态调整权限、必要时主动隔离(Nadella 强调"安全控制必须置于模型之外"——outside the model)。这一次升级的本质,是把 AI 代理从"被信任的工具"重新定义为"必须被审计的主体",也意味着 enterprise AI 治理正式进入"零信任 + 代理监督代理"双轨阶段。 参考来源: 1. Securing the agentic insider — 微软官方博客(2026-10-19) https://blogs.microsoft.com/blog/2026/10/19/securing-the-agentic-insider/ 2. Agents are your newest insiders. Security must move at their speed — 微软安全博客(2026-10-19) https://www.microsoft.com/en-us/security/blog/2026/10/19/agents-are-your-newest-insiders-security-must-move-at-their-speed/ 3. Microsoft says AI agents are like insider threats as 'frontier models can't be trusted' — TechCrunch(2026-10-19) https://techcrunch.com/2026/10/19/microsoft-says-ai-agents-are-like-insider-threats-as-frontier-models-cant-be-trusted/ 4. Microsoft CEO Satya Nadella says all AI agents are an 'insider risk' — NBC News(2026-10-19) https://www.nbcnews.com/tech/tech-news/microsoft-ceo-satya-nadella-says-ai-agents-are-insider-risk-rcna602355 5. Microsoft unveils tools to monitor and corral rogue AI agents — GeekWire(2026-10-19) https://www.geekwire.com/2026/microsoft-unveils-tools-to-monitor-and-corral-rogue-ai-agents/ 6. Microsoft Expands Project Perception to Address Security Risks from AI Agents — Redmondmag(2026-10-19) https://redmondmag.com/articles/2026/10/19/microsoft-expands-project-perception-to-address-security-risks-from-ai-agents.aspx 7. 'Frontier models cannot be trusted': Microsoft expands Project Perception — Firstpost(2026-10-19) https://www.firstpost.com/tech/frontier-models-cannot-be-trusted-microsoft-expands-project-perception-to-continuously-assess-ai-agent-risk-14053929.html 8. Microsoft: "Frontier AI models, regardless of their origin, cannot be trusted" — Neowin(2026-10-19) https://www.neowin.net/news/microsoft-frontier-ai-models-regardless-of-their-origin-cannot-be-trusted/ 9. Microsoft aims to secure agentic insiders with Project Perception — Techzine(2026-10-19) https://www.techzine.eu/news/security/145092/microsoft-aims-to-secure-agentic-insiders-with-project-perception/ 10. Satya Nadella: Securing the agentic insider — LinkedIn(2026-10-19) https://www.linkedin.com/posts/satyanadella_securing-the-agentic-insider-the-official-activity-7517971322266640384-ahKI 11. Microsoft Project Perception: agentic security system 概览 — Microsoft Security(2026) https://www.microsoft.com/en-us/security/business/ai-powered-cybersecurity/project-perception-agentic-system 12. What is Project Perception? — Microsoft Learn(2026) https://learn.microsoft.com/en-us/security/agentic-security/agentic-security-overview 13. Microsoft announces tools to treat AI agents as insider threats — Silicon Report(2026-10-19) https://www.siliconreport.com/microsoft-announces-tools-to-treat-ai-agents-as-insider-threats 14. Microsoft Securing the Agentic Insider: Project Perception & Guardian Agents — Explainx(2026-10-19) https://explainx.ai/blog/microsoft-securing-the-agentic-insider-project-perception-guardian-agents-2026 15. Microsoft just expanded Project Perception to treat AI agents as insider threats — AGTP(2026-10-19) https://x.com/AGTPinsights/status/2112200017939501447