字号·· | 护眼
CNBC

印度的AI风险:代理失控运行,安全团队疲于追赶India’s AI risks: Agents running unchecked, security teams playing catch-up

点「原文对照」整页切到原文,或双击某段只看那段的原文。

印度组织可能越来越习惯于部署人工智能代理,但它们确保这些系统安全和进行治理的能力却难以跟上步伐。ServiceNow的一项最新研究发现,该国92%的组织在过去一年中经历过人工智能代理安全事件。然而,受影响的组织中只有36%能够在24小时内遏制事件。

Indian organisations may be growing increasingly comfortable with deploying artificial intelligence (AI) agents, but their ability to secure and govern these systems is struggling to keep pace. A new ServiceNow study finds that 92% of organisations in the country experienced an AI agent security incident in the past year. Yet, only 36% of those affected were able to contain an incident within 24 hours.

这些调查结果表明,人工智能代理的采用与管理这些代理所需的的安全框架之间存在着日益扩大的差距。这份名为《部署但缺乏治理:印度代理安全现状》(Deployed but Ungoverned: The State of Agentic Security in India)的报告调查了该国的500名安全负责人,发现对代理安全的信心往往超出了实际到位的控制措施。

The findings point to a widening gap between the adoption of AI agents and the security frameworks needed to manage them. The report, Deployed but Ungoverned: The State of Agentic Security in India, surveyed 500 security leaders in the country and found that confidence in agentic security often runs ahead of the controls actually in place.

接受调查的安全负责人中有半数将其组织的安全代理描述为“已建立”。然而,只有19%的人表示有某位具名个人对人工智能代理安全治理负责,而只有18%的人对其代理拥有统一的、单一界面的可视性。

Half of the security leaders surveyed described their organisation's agentic security as "established". However, only 19% said a single named individual was accountable for AI agent security governance, while just 18% had unified, single-interface visibility into their agents.

这些差距延伸到了部署和合规方面。只有30%的组织拥有自动化部署关卡,能够在人工智能代理通过安全检查之前阻止它们。不到一半(即47%)的组织表示它们可以按需提供完整的证据包。

The gaps extend to deployment and compliance. Only 30% of organisations have an automated deployment gate that blocks AI agents until they clear security checks. Fewer than half, or 47%, said they could produce a complete evidence package on demand.

人工智能代理现在已成为印度企业运营方式的一部分,但安全性并未跟上步伐 > 库尔米特·巴瓦(KULMEET BAWA),SERVICENOW印度及南亚区域(SAARC)全球副总裁兼董事总经理。

AI agents are now part of how Indian enterprises operate, but security has not kept pace > KULMEET BAWA, GVP AND MD FOR INDIA AND SAARC, SERVICENOW.

他表示:“你无法保护或治理你看不见的东西。”他呼吁将每个代理进行注册、指派所有者并在一个地方实现可视化,同时自动实施控制措施。

"You can’t secure or govern what you can’t see," he said, calling for every agent to be registered, assigned an owner and made visible in one place, with controls enforced automatically.

检测领先于响应 该研究强调了安全链中的另一个薄弱环节:组织可能在识别可疑行为方面变得更好了,但许多组织在没有人工干预的情况下仍然难以做出响应。

Detection is ahead of response The study highlights another weak link in the security chain: organisations may be getting better at spotting suspicious behaviour, but many still struggle to respond without human intervention.

约 51% 的印度组织已实现对异常 AI 智能体行为的自动化实时检测。然而,67% 的组织无法在智能体出现异常行为时自动采取行动。这在识别潜在威胁与遏制威胁之间留下了空档,尤其是在 AI 智能体能够持续且以机器速度运行的情况下。

About 51% of Indian organisations have automated real-time detection for anomalous AI agent behaviour. However, 67% cannot automatically act when an agent behaves anomalously. This leaves a gap between identifying a potential threat and containing it, particularly when AI agents can operate continuously and at machine speed.

安全团队介入流程的时机也过晚。仅 46% 的组织在 AI 智能体开发开始前让安全团队参与,这意味着控制措施可能未从一开始就内嵌到系统中。

Security teams are also being brought into the process too late. Only 46% of organisations involve security teams before AI agent development begins, meaning controls may not be built into systems from the outset.

风险并不局限于传统 IT 环境。报告发现,83% 的组织拥有在运营技术(OT)和物理环境中运行、或与之交互的 AI 智能体。然而,仅 29% 的组织拥有专为这些场景设计的独立治理框架。

The risks are not limited to conventional IT environments. The report found that 83% of organisations have AI agents operating in, or interacting with, operational technology (OT) and physical environments. Yet, only 29% have a separate governance framework designed for these settings.

未登记的 AI 智能体增加了风险 所谓的“影子 AI”蔓延是另一大担忧。 约 37% 的组织在过去 12 个月内发现生产环境中运行着从未向其安全团队登记、也未为安全团队所知的 AI 智能体。此类系统可能成为盲区,使组织对其网络内运行的事物缺乏有效监管。

Unregistered AI agents add to the risk The spread of so-called "Shadow AI" is another concern. Around 37% of organisations discovered AI agents running in production over the past 12 months that had never been registered with, or known to, their security teams. Such systems can become blind spots, leaving organisations with limited oversight of what is operating within their networks.

随着监管机构和网络安全机构将关注点转向 AI 智能体带来的风险,上述调查结果随之出台。CERT-In(印度计算机应急响应小组)2026 年 10 月发布的关于 AI 智能体的通报,已将可见性与控制问题推至聚光灯下。此外,根据 CERT-In 2022 年发布的指令,印度组织须在知悉网络安全事件后六小时内上报。

The findings come as regulators and cybersecurity agencies turn their attention to the risks posed by AI agents. CERT-In's October 2026 advisory on AI agents has brought the issue of visibility and control into focus. Separately, under CERT-In's 2022 directions, organisations in India are required to report cyber incidents within six hours of becoming aware of them.

研究还发现,使用统一智能体安全平台的组织报告称拥有更强的能力。在这些受访者中,62% 具备针对异常智能体行为的自动化实时警报,而未使用此类平台的组织这一比例为 43%。自动化修复策略违规的相应比例分别为 45% 和 24%。

The study also found that organisations using a unified platform for agentic security reported stronger capabilities. Among these respondents, 62% had automated real-time alerts for unusual agent behaviour, compared with 43% of organisations not using such platforms. The corresponding figures for automated remediation of policy violations were 45% and 24%, respectively.

研究传递的信息很明确:部署 AI 智能体只是方程式的一部分。组织还需要了解有哪些智能体在运行、谁对它们负责,以及安全控制能否像这些系统本身一样快速做出反应。

The message from the research is clear: deploying AI agents is only one part of the equation. Organisations also need to know which agents are running, who is responsible for them and whether security controls can respond as quickly as the systems themselves.

该研究由 ServiceNow 与 Phronesis Partners 联合开展。调查对象为 20 个国家的 3700 多名副总裁级及以上的高级安全和技术决策者,其中包括印度的 500 名安全负责人。全球研究共纳入亚太地区 1201 名受访者。印度的 AI 风险:智能体不受控运行,安全团队疲于奔命

The study was conducted by ServiceNow in partnership with Phronesis Partners. It surveyed more than 3,700 senior security and technology decision-makers at vice-president level and above across 20 countries, including 500 security leaders in India. The global study included 1,201 respondents across the Asia-Pacific region. India’s AI risks: Agents running unchecked, security teams playing catch-up

阅读媒体原文 ↗