网络安全
CrowdStrike 新增 AI 向猎鹰守护者切换Rogue Agents
在 Google 上的首选来源中添加 Unite.AI
CrowdStrike 宣布 2026年11月3日为猎鹰卫报的AI Kill Switch，为安全小组提供了人控制的方式来打断流氓AI agents. 该功能可在发布获得，并使用连接的AI网关拒绝agents访问AI模型。
发布解决了agents，它可以连锁任务，并用系统级接入在端点，云环境和SaaS应用中运行。 CrowdStrike 称这些agents可以遇到障碍，采取无人授权的行动，即使没有涉及对手。 杀机开关需要经授权的管理员触发响应，每个动作都记录在审计线索中。
网关执行是当前控制
执法是通过一般可用的猎鹰AI网关或已经使用的第三方AI网关进行的。 CrowdStrike网关与Guardian一起不增加费用。 它巩固了猎鹰平台的模型访问，安全政策，证书管理，以及使用和成本控制。 CrowdStrike 称网关还将保护范围扩大到直接端点感应范围以外的应用程序和工作量，包括未经管理的环境。
"卫报"产品文档将当前响应定义为削减访问指定的AI模型和提供者的集中执行，而由agent类型标记的网关执行即将到来。 因此，直接控制涉及agent与AI服务的联系，这个机制比名称可能更狭窄。
守护者用品活动背景
Falcon Guardian本身自2026年9月起普遍可供使用，因此杀机开关延长了现有的产品。 守护者已经发现企业AI agents，对其行动执行政策并发现针对他们的攻击。 CrowdStrike Agent Graph将agent的指示与后续执行以及所涉及的身份，工具，系统和资源联系起来，使调查人员能够追踪活动和影响。
CrowdStrike将这些控制点合并的理由是，agent级和仅限网络的视图会留下不同的盲点。 公司称 agent级钩子可以看到提示，但可以错过模型响应和一些动作，而只使用网络的工具则无法观察活动从未穿越网络的主机执行或本地和开放量级模型。 守护者将agent钩子，端点网络检查和端点运行时间在一只猎鹰传感器上的可见度结合在一起。 端点网络检查一般为Windows，早期为Mac访问。
路线图上的更大限制
发布公告给出了更广泛的回应：终止agent运行的地方，阻止目标主机或主机组重启，并取消其企业网络访问。 它本身的可用序列将终止放在了网关上，并很快放在了终点上。 直接终止终点agents及其相关进程，以及防止重启，是即将到来的网关强制执行之外的扩展的一部分。
计划的协调封装还包括终止或包含云 agents，运行时取消agent权限，限制或终止SaaS环境中的agent活动，以及终止浏览器会话。 Charlotte Agentic SOAR意在协调这些执行点，并启用由AI特定检测或风险agent行为引发的自动杀开关工作流程。 CrowdStrike 称扩展旨在针对单个agents或组别，同时尽量减少对更广泛业务的干扰。
发布分离了两个问题，这两个术语的杀换可以模糊：安全小组是否能够立即拒绝agent的模型连接，以及他们是否能够在agent运行的地方禁用agent的流程和权限。 航运管制通过网关解决第一个问题；第二个问题取决于CrowdStrike路线图上仍然保持的协调执法。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Cybersecurity
CrowdStrike Adds AI Kill Switch to Falcon Guardian for Rogue Agents
Add Unite.AI to your preferred sources on Google
CrowdStrike announced the AI Kill Switch for Falcon Guardian on November 3, 2026, giving security teams a human-controlled way to interrupt rogue AI agents. The feature is available at launch and uses connected AI gateways to deny agents access to AI models.
The launch addresses agents that can chain tasks and operate with system-level access across endpoints, cloud environments and SaaS applications. CrowdStrike says these agents can encounter obstacles and take actions no one authorized, even when no adversary is involved. The kill switch requires an authorized administrator to trigger the response, and every action is recorded in an audit trail.
Gateway Enforcement Is the Current Control
Enforcement runs through either the generally available Falcon AI Gateway or a third-party AI gateway already in use. The CrowdStrike gateway is included with Guardian at no additional cost. It consolidates model access, security policies, credential management, and usage and cost controls in the Falcon platform. CrowdStrike says the gateway also extends protection to applications and workloads outside direct endpoint-sensor coverage, including unmanaged environments.
The Guardian product documentation defines the current response as centralized enforcement that cuts access to designated AI models and providers, with gateway enforcement by agent type marked as coming soon. The immediate control therefore concerns the agent’s connection to AI services, a narrower mechanism than the name might suggest.
Guardian Supplies Activity Context
Falcon Guardian itself has been generally available since September 2026, so the kill switch extends an existing product. Guardian already discovers enterprise AI agents, enforces policies on their actions and detects attacks against them. CrowdStrike Agent Graph links an agent’s instructions to subsequent execution and the identities, tools, systems and resources involved, allowing investigators to trace both activity and impact.
CrowdStrike’s rationale for combining these control points is that agent-level and network-only views leave different blind spots. The company says agent-level hooks see prompts but can miss model responses and some actions, while network-only tools cannot observe host execution or local and open-weight models whose activity never crosses the network. Guardian combines agent hooks, endpoint network inspection and endpoint runtime visibility on one Falcon sensor. Endpoint network inspection is generally available for Windows and in early access for Mac.
Wider Containment Remains on the Roadmap
The launch announcement presents a broader response: terminating an agent where it runs, blocking relaunches on a targeted host or host group, and removing its enterprise network access. Its own availability sequence puts termination first at the gateway and soon on the endpoint. Direct termination of endpoint agents and their associated processes, as well as prevention of restarts, are part of the coming-soon expansion beyond gateway enforcement.
Planned coordinated containment also covers terminating or containing cloud agents, revoking agent permissions at runtime, restricting or ending agent activity in SaaS environments, and terminating browser sessions. Charlotte Agentic SOAR is intended to coordinate those enforcement points and enable automated kill-switch workflows triggered by AI-specific detections or risky agent behavior. CrowdStrike says the expansion is designed to target individual agents or groups while minimizing disruption to broader business operations.
The launch separates two questions that the term kill switch can blur: whether security teams can immediately deny an agent its model connection, and whether they can disable the agent’s processes and permissions wherever it operates. The shipping control addresses the first through gateways; the second depends on coordinated enforcement that remains on CrowdStrike’s roadmap.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.