网络安全
CrowdStrike 目标 Rogue AI Agents 内幕威胁探测器
在 Google 上的首选来源中添加 Unite.AI
2026年11月4日，CrowdStrike 宣布在巴塞罗那的Fal.Con Europe，将在Q4中将Agent内幕威胁探测器加入猎鹰Next-Gen SIEM. 公告将即将具备的能力与一般可协调的多agent调查联系起来，将行为警报与跨领域调查和反应联系起来。
目标案例是拥有系统级访问敏感数据以及采取行动的自主权的agent企业。 如果agent在不违反规则的情况下被操纵或偏离原定行为，基于政策和日志的工具就没有明显的违反行为可捕。 他们也可以努力区分agent的行为与运行者的行为，留下一个行为问题，在单个事件层面看起来是合法的。
Agent 内幕威胁探测(Insider President Detection)旨在为每个AI agent及其所有者学习正常行为，然后即使没有违反政策也会出现国旗偏差。 它的新增对计分猎鹰Next-Gen SIEM为主人和主机构建的风险，并将异常行为直接引入协调调查。 因此，触发因素超越了破解的规则，转向行为改变，身份和主机背景都包含在调查中。
协调调查普及
调查现在一般是2026年9月2日在法尔的发布。 在拉斯维加斯 Charlotte AI在端点，身份，SaaS，云域和网络域平行发送专家训练的agents. 他们在一个共同的背景层上工作，而不是进行孤立的调查，然后以显而易见的推理和理由就裁决达成一致。
这一设计解决了跨领域调查中的组装问题。 分析人员可以检查agents的推理，而不是从不同的领域手动缝合结果，而共享上下文则允许一个agent学到的信息向从事同一调查的其他人通报。
检测和调查为包括猎鹰守护者AI Kill Switch在内的一个响应层提供素材，它终止了agent. 分析员还有Agent中断暂停或停止运行的夏洛特AI或AgentWorks agent中任务，这种控制旨在防止未经授权的行动和逃逸支出。
人类控制和下一个探测阶段
Charlotte Agentic SOAR作为agents，工作流程和自定义应用程序的单一规范工作空间可供使用，自主设置从人类批准到完全自主执行不等。 当它引用敏感工具并提出行动请求时，它的人入室门暂停agent. 分析员可以批准行动，有文件证明的理由拒绝行动，或者检查agent的全部痕迹，围绕需要更严密控制的工具建立一个可审查的决定点。
同一Q4路线图列出了若干相关控制。 AI生态系统检测旨在涵盖LLM API滥用，被盗API密钥，以及涉及编码助理，MCP服务器和agent框架的妥协。 核准闸机计划从单个工具呼叫扩展到整个agents，应用程序和工作流程。
AI Credit Visicity and Control旨在提供实时消费能见度，中途停运，以及当agent达到上限时决定是否提高或拒绝限制。 这就把开支控制与安全批准放在一起，使管理人员在资源消耗变得不可接受时有办法中断自主工作。
后期项目"政策分裂Agent Guardrails"定于2028年财政第一季度。 Falcon Guardian将筛选Charlotte AI和Agent Works的相互作用，以便进行政策漂移、未经授权的工具呼叫和出界数据访问。 未释放的服务和特征仍在开发之中，可能会发生变化，使这些标签和时间表成为路线图，而不是目前的可用性。
该版分几个阶段组装agent-安全工作流程：现在已有协调的调查和操作员控制，为Q4列出行为检测，政策驱动护栏更靠前。 对维权者来说，这给审查或阻止agent的行动和确定应该引发反应的微妙行为变化留下了实际的区别。
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 Targets Rogue AI Agents With Insider Threat Detection
Add Unite.AI to your preferred sources on Google
CrowdStrike announced on November 4, 2026, at Fal.Con Europe in Barcelona that it will add Agent Insider Threat Detection to Falcon Next-Gen SIEM in Q4. The announcement pairs that forthcoming capability with generally available coordinated multi-agent investigations, connecting behavioral alerts to cross-domain investigation and response.
The target case is an enterprise agent with system-level access to sensitive data and the autonomy to act on it. If that agent is manipulated or drifts from its intended behavior without violating a rule, policy- and log-based tools have no explicit violation to catch. They can also struggle to distinguish the agent’s actions from those of the person running it, leaving a behavioral problem that can appear legitimate at the level of individual events.
Agent Insider Threat Detection is designed to learn normal behavior for each AI agent and its owner, then flag deviations even when no policy is violated. It adds that risk to the score Falcon Next-Gen SIEM builds for the owner and host, and routes abnormal behavior directly into a coordinated investigation. The trigger therefore moves beyond a broken rule to a change in behavior, with identity and host context carried into the investigation.
Coordinated Investigations Reach General Availability
The investigations now generally available were unveiled on September 2, 2026, at Fal.Con in Las Vegas. Charlotte AI dispatches expert-trained agents in parallel across endpoint, identity, SaaS, cloud and network domains. They work on a shared context layer rather than conducting isolated inquiries, then converge on a verdict with visible reasoning and justification.
That design addresses the assembly problem in cross-domain investigations. Analysts can examine the agents’ reasoning instead of manually stitching together findings from separate domains, while shared context allows information learned by one agent to inform the others working on the same investigation.
Detection and investigation feed a response layer that includes the Falcon Guardian AI Kill Switch, which terminates an agent. Analysts also have Agent Interruption to pause or stop a running Charlotte AI or AgentWorks agent mid-task, a control intended to prevent unauthorized action and runaway spending.
Human Controls and the Next Detection Phase
Charlotte Agentic SOAR is available as a single governed workspace for agents, workflows and custom applications, with autonomy settings ranging from human approval to fully autonomous execution. Its human-in-the-loop gates pause an agent when it invokes a sensitive tool and raise an Action Request. Analysts can approve the action, reject it with a documented reason or inspect the full agent trace, creating a reviewable decision point around tools that warrant closer control.
Several related controls are listed for the same Q4 roadmap. AI Ecosystem Detections are intended to cover LLM API abuse, stolen API keys, and compromises involving coding assistants, MCP servers and agent frameworks. Approval gates are planned to extend from individual tool calls to entire agents, applications and workflows.
AI Credit Visibility and Control is designed to provide real-time consumption visibility, mid-run halts and decisions on whether to raise or deny limits when an agent reaches its cap. That places spending controls alongside security approvals, giving administrators a way to interrupt autonomous work when its resource consumption becomes unacceptable.
A later item, Policy-Drift Agent Guardrails, is scheduled for the first quarter of fiscal 2028. Falcon Guardian would screen Charlotte AI and AgentWorks interactions for policy drift, unauthorized tool calls and out-of-bounds data access. The unreleased services and features remain in development and subject to change, making these labels and schedules a roadmap rather than current availability.
The release assembles an agent-security workflow in stages: coordinated investigation and operator controls are available now, behavioral detection is listed for Q4, and policy-drift guardrails sit further out. For defenders, that leaves a practical distinction between reviewing or stopping an agent’s action and identifying the subtle behavioral change that should trigger the response.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.