Sysdig 发布 AI Defense to Govern Agent 从终点到云端的动作

网络安全

Sysdig 发布 AI Defense to Govern Agent 从终点到云端的动作

在 Google 上的首选来源中添加 Unite.AI

Sysdig拥有发布 AI Defense，这是一个运行时间安全和治理产品，旨在检查AI agents做什么，并在他们的行为超过一个组织的政策时进行干预。 11月4日的公告使得该产品可以普遍用于macOS和Windows，其能见度旨在将员工工作站的活动与他们能达到的云资源联系起来。

该产品涉及从生成答案的AI到操作软件的AI的转变。 一个编码agent可以读取文件，引用工具，启动进程和更改基础设施。 要保证工作流程，既需要了解agent认为它正在执行的任务，也需要了解它下面实际发生的操作。

连接 Agent 意图与操作系统证据

Sysdig 称 AI Defense将agent内置的仪器——运行它的框架——与操作系统遥测相结合。 利用仪器捕捉到对工具的呼叫，模型背景协议服务器，技能和插件，包括他们的论据和目标。 系统层面的观察单独记录这些调用产生的过程和活动。

当agent的账户不完整或受损时，区别就很重要。 一个报告的工具动作可以启动几个进程，一个会话中添加的连接或技能可以改变agent的能力。 Sysdig报告说，对agent日志中每个动作的大约7个基本过程进行了早期观察；这是公司发现，而不是所有agents的普遍比率。

该系统还考虑权威：同一文件读取操作可产生不同的后果，取决于所涉及的证书是到达测试环境还是管理生产。 行动图旨在将事件序列、责任人、agent申报的活动和可获取的资源联系起来。

政策允许、暂停或屏蔽动作

各组织可以在实施之前以警戒模式启动政策。 一旦强制执行，Sysdig 称这些政策可以阻止未经审查的工具和连接，停止冒险的认证访问，或采取得到人类批准的行动。 发布中还描述了如果一个动作超越了先前的控制，则终止一个过程。

这些机制在执行过程中服务于不同的点。 防止行动完成不同于发现活动并终止其后的进程。 该公告没有提供一个公共矩阵，显示每个支持的操作系统、使用和管理的平台上有哪些控制是预防性的。

政策模板包括用于LLM应用的OWASP Top 10的映射，但开发者除外路径。 Sysdig还描述了会话记录，显示适用了哪项政策，允许或被屏蔽了什么，以及谁批准了例外。 这些记录可以支持调查和审计工作，而不需要各小组在事件发生后重建会议。

法尔科基金会新增

Sysdig将产品与其在云母运行时安全以及Falco，它最初创建的开源项目的历史联系起来。 Falco的文档描述了观察Linux内核事件和插件提供的事件源，用容器和Kubernetes上下文丰富事件，并应用规则生成警报。

这个基础有助于解释为什么Sysdig强调观察处决而不是仅仅依靠AI系统的叙述。 这并不意味着开源 Falco项目本身为AI Defense中描述的每个agent治理或阻塞能力提供补给。 Falco记录的检测架构和Sysdig的商业产品应该分别理解。

在agent环境中的发布名称为Claude Code,Codex,GitHub Copilot和Cursor，旨在盘点并追踪人类所有者。 其macOS和Windows可用性说明比延伸至Linux,Kubernetes和管理的agent平台的更广泛的覆盖范围描述更为具体；各组织需要建立适用于每一种环境的整合。

运行时治理补充模型测试

Sysdig关于共享责任框架运行时间控制的技术讨论，实际的区别是要求一个模型尊重边界，并在软件试图操作时强制执行边界。

AI Defense与现有的端点保护，模型测试和红色组合并列。 测试检查一个模型或工作流程在选定条件下的表现；运行时治理处理实际会话中的行动。 两项说明都没有规定部署不得妥协，效力取决于政策配置和执行路径的涵盖范围。

分配 Agent 活动和 AI 开支

除了安全之外，Sysdig 称 AI Defense Companys还用agent，团队和负责人调用，包括不能通过中央AI网关的活动。 这有助于各组织区分日常工作中昂贵的模型用途和需要更有能力的模式的活动。

更广泛的主张是安全、治理和成本管理的共同证据层。 随着agents获得更多工具，并在真正的员工证书下运作，其影响由他们能够执行什么来定义。 爱思潮辩护的目的是使该权威在行动点能够被看到和强制执行。

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

Sysdig Launches AI Defense to Govern Agent Actions From Endpoint to Cloud

Add Unite.AI to your preferred sources on Google

Sysdig has launched AI Defense, a runtime security and governance product designed to inspect what AI agents do and intervene when their actions exceed an organization’s policies. The November 4 announcement makes the product generally available for macOS and Windows, with visibility intended to connect activity on employee workstations to the cloud resources their credentials can reach.

The product addresses the shift from AI that produces answers to AI that operates software. A coding agent can read files, invoke tools, start processes and change infrastructure. Securing that workflow requires understanding both the task the agent believes it is performing and the operations actually occurring underneath it.

Connecting Agent Intent With Operating-System Evidence

Sysdig says AI Defense combines instrumentation inside an agent’s harness—the framework that runs it—with operating-system telemetry. Harness instrumentation captures calls to tools, Model Context Protocol servers, skills and plugins, including their arguments and targets. System-level observation provides a separate record of the processes and activity those calls produce.

The distinction matters when an agent’s account is incomplete or compromised. One reported tool action can start several processes, and a connection or skill added during a session can change what the agent is capable of doing. Sysdig reports an early observation of roughly seven underlying processes for each action in agent logs; that is a company finding, rather than a universal ratio for all agents.

The system also considers authority: the same file-reading operation can carry different consequences depending on whether the credential involved reaches a test environment or administers production. An action graph is intended to connect the sequence of events, the accountable person, the agent’s stated activity and the resources within reach.

Policies That Allow, Pause or Block an Action

Organizations can start policies in alert mode before enabling enforcement. Once enforced, Sysdig says those policies can block unvetted tools and connections, stop risky credential access or hold an action for human approval. The release also describes terminating a process if an action gets past an earlier control.

These mechanisms serve different points in an execution sequence. Preventing an operation from completing is distinct from detecting activity and terminating a process afterward. The announcement does not provide a public matrix showing which controls are preventive on each supported operating system, harness and managed platform.

Policy templates include mappings to the OWASP Top 10 for LLM Applications, with developer exception paths. Sysdig also describes session records showing which policy applied, what was allowed or blocked and who approved an exception. Those records can support investigation and audit work without requiring teams to reconstruct a session after an incident.

What the Falco Foundation Adds

Sysdig connects the product to its history in cloud-native runtime security and Falco, the open-source project it originally created. Falco’s documentation describes observing Linux kernel events and plugin-provided event sources, enriching events with container and Kubernetes context, and applying rules to generate alerts.

That foundation helps explain why Sysdig emphasizes observed execution rather than relying solely on an AI system’s narrative. It does not mean the open-source Falco project itself supplies every agent-governance or blocking capability described in AI Defense. Falco’s documented detection architecture and Sysdig’s commercial product should be understood separately.

The release names Claude Code, Codex, GitHub Copilot and Cursor among the agent environments it aims to inventory and trace to a human owner. Its macOS and Windows availability statement is more specific than the broader description of coverage extending into Linux, Kubernetes and managed agent platforms; organizations will need to establish the applicable integration for each environment.

Runtime Governance Complements Model Testing

Sysdig’s technical discussion of shared responsibility frames runtime controls as a layer around the work an agent performs. The practical distinction is between asking a model to respect a boundary and enforcing that boundary when software attempts an operation.

AI Defense is positioned alongside existing endpoint protection, model testing and red teaming. Testing examines how a model or workflow behaves under selected conditions; runtime governance addresses actions during an actual session. Neither description establishes that a deployment is immune to compromise, and effectiveness depends on policy configuration and coverage of the execution path.

Attributing Agent Activity and AI Spending

Beyond security, Sysdig says AI Defense associates model calls with the agent, team and person responsible, including activity that does not pass through a centralized AI gateway. That could help organizations distinguish expensive model use on routine work from activity that warrants a more capable model.

The broader proposition is a common evidence layer for security, governance and cost management. As agents acquire more tools and operate under real employee credentials, their impact is defined by what they can execute. Sysdig AI Defense aims to make that authority visible and enforceable at the point of action.

Auto-fetched and translated by AI Atlas. All rights belong to the original author.