AI 将在 2027 年如何改变企业防御的 5 种方式

NiPlot/ iStock / Getty Images Plus via Getty Images
Google 预计,到 2027 年,智能体式 AI 将大规模影响安全领域。
安全专家表示,AI 的响应速度超过纯人工响应团队,已成为我们的新现实。
智能体需要一个身份,并且必须作为资产和网络安全风险来管理。
今年年初,很少有组织在其网络安全工作流中使用大语言模型(LLM)和 AI 智能体,包括用于发现和修补漏洞。据 Google 称,如今这种情况已在大规模发生。
智能体式 AI 的能力正迅速超越传统安全运营中心(SOC)所能运转的速度和规模。无论你是否愿意,AI 在新现实中都扮演着不可或缺的角色——对防御方和威胁行为者皆然。
速度和规模是企业考虑在 SOC 以及更广泛的安全领域采用 AI 的主要原因,这些因素也支撑了 Google 针对 2027 年的最新安全预测——Cloud CISO Perspectives。然而,若没有恰当的预防措施,采用 AI 的安全实践者可能无意中扩大而非缩小其攻击面。
1. SOC 中的威胁警报逐级上报
告警疲劳在 SOC 团队中早已是众所周知的问题。一条接一条的告警需要逐条翻查;其中许多无关紧要或是误报,让高技能人员把时间耗在繁琐、枯燥且耗时的告警分诊工作上。
如今,自主的 AI 驱动型威胁可以在几分钟内探测网络、获取入口并在其中横向移动,因此再去追查来自多个系统、记录于数天甚至数周前的威胁数据和安全告警,已无法帮助企业缓解这些威胁——更不用说跟上威胁的速度了。
根据 Google 的预测,AI 将意味着告警队列悄然消失。Google Cloud Security 工程副总裁 Ramya Chitrakar 认为,“2027 年将是分析师不再追逐告警、转而塑造结果的一年。”
如何实现?CISO 必须改变告警模型本身。按照 Google 的说法,不再由人工防御者在网络安全事件发生之初对告警进行分诊,而是让告警成为最终输出,由自主智能体“自动完成数据丰富、关联与分诊,呈现经过验证的结论,而不是一条原始工单”。
这意味着网络安全专家可以监督告警处理并核验自动化操作,从而有可能腾出精力专注于其他任务。
2. 智能体式 AI 需要身份标识与审计留痕
定制化的智能体式 AI 往往被设计成——或至少意图是——在人工监督下自主工作。然而,成千上万个独立智能体在多种环境中运行,可能在企业网络中制造出被网络攻击者盯上的高价值入口。
我们需要把智能体式 AI 视为不止是一种工具:每一个智能体都需要像人类员工一样拥有自己的身份,以便在零信任架构内运行、保持控制权并降低被攻陷的风险。
不过,谷歌安全工程副总裁 Heather Adkins 认为这还远远不够,她提出零信任理念不应止步于网络层。Adkins 认为,它需要延伸到交易层。
必须为每个智能体建立可追溯的记录,而这靠人工根本无法处理;即便在零信任框架内,AI 智能体也可能被赋予过大的权限范围。正确的做法是:对每一笔交易和每一次授权都实施验证,并为每个 AI 智能体分配各自的最小权限凭证,一旦发生安全事件或该智能体不再被需要,就可以将其撤销。
报告指出:“在实时自主推理的支撑下,我们可以动态放行周二的一笔交易,而如果周三条件看起来异常,就能拦截完全相同的请求。”
3. 软件开发速度加快
AI 正被用于加快新软件的开发和部署,而谷歌认为,到 2027 年,安全必须成为一个嵌入软件交付流水线的持续运行系统。
多年来,威胁行为者一直在实施软件供应链攻击,其中最引人注目的当属 2020 年的 SolarWinds 事件。谷歌威胁情报副总裁 Sandra Joyce 表示,网络犯罪分子如今正在污染从开源代码到软件库的一切,而 AI 编程助手往往会自动把这些内容拉取进代码库。
响应需要数天的缺陷修复工单,以及安全只出现在软件开发周期末端的做法,都已不再可行。相反,编写代码与保障代码安全必须同时进行,并借助智能体式 AI,让速度不以牺牲安全为代价。
“同样至关重要的是,我们要把实时生产环境直接关联回源代码,这样防御者就能把现实世界中的暴露面追溯到确切的代码行,验证它是否真的可被利用,并在不破坏生产环境的前提下,以机器速度交付经过预先测试的修复方案,”Google 补充道。
4. 对智能体式 AI 开展红队测试
当各组织争相把 AI 引入其应用和服务,以显得“AI 优先”、富有创新且紧跟潮流时,安全可能被忽视,或被当成事后才考虑的事。
因此,我们面对的攻击面可能比我们意识到的更新、更广,因为 AI 带来了新的安全问题,包括软件漏洞、错误配置、绕过手段以及间接提示注入攻击。如今 AI 也掌握在所有网络攻击者手中——从脚本小子到国家支持的组织——因此审计企业网络和基础设施至关重要。
渗透测试和红队演练现在需要覆盖智能体式 AI。这在几年前还不在许多红队的范围内,但如今企业已热衷于在工作流、流程和产品中落地 AI,组织需要在威胁行为者加以利用之前,识别并修复 AI 引入的安全问题。
CISO 应指派渗透测试人员在自主工作流上线生产之前进行审计,然后将结果与威胁情报结合,确保防御者掌握所需信息,以加强网络防御,并针对新型和前所未有的攻击加固分类器。
我们可能在 2027 年看到的最后一个重要趋势,是治理智能体泛滥问题,打造能够大规模快速处理安全任务的协同 AI 智能体团队。
当 AI 辅助的攻击者能够在比以往短得多的时间内攻陷终端设备、进入网络、横向移动或开始窃取数据时,仅靠获得通用指令的单个 AI 智能体已经不够用了。CISO 应当用专业化、协同编排的 AI 智能体团队来取代这些“通才型”AI 智能体,这些团队能够借助专业工具和技能组合,满足防御 AI 驱动攻击的需求。
CISO 还需要确保每个智能体都拥有各自的领域训练与理解、记忆和专用工具,并能在既定界限内自主行动,同时由一个中央协调者判断何时把决策交给人类。这需要监督,但作为回报,组织将避免智能体泛滥——这家科技巨头认为,这是影子 IT 及其相关威胁的下一次迭代。
Charlie Osborne 是一名网络安全记者兼摄影师,在伦敦为 ZDNET 和 CNET 撰稿。PGP 密钥:AF40821B 查看完整简介
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(5 ways AI will change how you defend your business in 2027)
NiPlot/ iStock / Getty Images Plus via Getty Images
Google expects agentic AI to influence security at scale in 2027.
Security experts say AI outpacing human-only response teams is our new reality.
Agents need an identity and must be managed as assets and cybersecurity risks.
At the beginning of the year, few organizations were utilizing large language models (LLMs) and AI agents in their cybersecurity workflows, including for discovering and patching vulnerabilities. According to Google, this is now happening at scale.
The capabilities of agentic AI are fast outstripping the speed and scale at which traditional security operations centers (SOCs) can operate. Like it or not, AI has an integral role in the new reality — for the defenders and threat actors.
Speed and scale are the main reasons enterprises consider adopting AI in SOCs and for security more broadly, and these factors underpin Google’s new security forecast for 2027 — Cloud CISO Perspectives. However, without the right precautions, AI security adopters could inadvertently increase their attack surface rather than shrink it.
1. Threat alerts in your SOC move up the chain
Alert fatigue is well-known across SOC teams. Alert after alert needs to be trawled through; many are irrelevant or false positives, tying up skilled staff with the mundane, tedious, time-consuming task of triaging them.
Now that autonomous, AI-driven threats can probe a network, gain entry, and move laterally across it in minutes, chasing threat data and security alerts from multiple systems logged days or weeks ago can’t help a business mitigate them — and certainly not at the speed of the threat.
According to Google’s forecast, AI will mean the alert queue quietly disappears. Ramya Chitrakar, VP of engineering at Google Cloud Security, believes that “2027 is the year analysts stop chasing alerts and start shaping outcomes.”
How? CISOs must change the alert model itself. Instead of human defenders triaging alerts at the start of a cybersecurity incident, the alert becomes final output, with autonomous agents handling “data enrichment, correlation, and triage automatically, surfacing a verified conclusion rather than a raw ticket,” according to Google.
This means cybersecurity experts can oversee alert handling and verify automated actions, potentially freeing them to focus on other tasks.
2. Agentic AI needs identities and paper trails
Custom agentic AI is often designed — or at least intended — to work autonomously, with human oversight. However, having thousands of individual agents working across multiple environments can create high-value entry points into corporate networks that cyberattackers target.
We need to think of agentic AI as more than a tool: each one needs its own identity, like a human employee, to operate within a zero-trust architecture, retain control, and reduce the risk of compromise.
However, Heather Adkins, VP of security engineering at Google, thinks this needs to go further and suggests the zero-trust concept shouldn’t stop at the network layer. Instead, Adkins believes it needs to reach the transaction layer.
A paper trail for each agent needs to be established, which is impossible to handle manually, and even within a zero-trust framework, AI agents could be given too much scope. Instead, implement verification for each transaction and for every authorization, and assign each AI agent its own least-privilege set of credentials that can be revoked if there is a security incident or the agent is no longer required.
“When backed by real-time autonomous reasoning, we can dynamically allow a transaction on Tuesday and block the exact same request on Wednesday if conditions look anomalous,” the report notes.
3. Software development speeds escalate
AI is being used to speed up the development and deployment of new software, and in 2027, security will have to become a continuous system embedded in the software delivery pipeline, according to Google.
Threat actors have carried out software supply chain attacks over the years, one of the most notable being the SolarWinds incident of 2020. Sandra Joyce, VP of Google Threat Intelligence, says cybercriminals are now poisoning everything from open-source code to software libraries, which AI coding assistants often pull automatically into codebases.
Bug-fix tickets that take days to respond to and security appearing at the end of a software development cycle no longer work. Instead, writing and securing code must happen simultaneously, with the help of agentic AI, so speed doesn’t compromise security.
“It’s also crucial that we link live production environments directly back to source code so that defenders can trace a real-world exposure to the exact line of code, validate whether it is genuinely exploitable, and deliver a pre-tested fix at machine speed without breaking production,” Google added.
4. Red-teaming agentic AI
As organizations rush to implement AI into their applications and services to appear “AI-first,” innovative, and modern, security may be ignored or treated as an afterthought.
As a result, we may face a new, broader attack surface than we realize, as AI introduces new security issues, including software vulnerabilities, misconfiguration, bypasses, and indirect prompt injection attacks. Now that AI is also in the hands of all cyberattackers — from script kiddies to state-sponsored groups — auditing corporate networks and infrastructure is crucial.
Penetration testing and red team exercises now need to cover agentic AI. This wasn’t in scope for many red teams only a few years ago, but now that businesses have run with the idea of implementing AI in their workflows, processes, and products, organizations need to identify and fix security issues introduced by AI before threat actors take advantage of them.
CISOs should task pen testers with auditing before an autonomous workflow hits production, then combine the results with threat intelligence to ensure defenders have the information they need to bolster network defenses and harden classifiers against new and novel attacks.
The final significant trend we are likely to see in 2027 is tackling agent sprawl to create coordinated teams of AI agents that can rapidly handle security tasks at scale.
Single AI agents given general instructions won’t cut it when AI-assisted attackers can compromise endpoint devices, enter networks, move laterally, or start stealing data in a far shorter timeframe than before. Instead, CISOs should replace these “generalist” AI agents with specialized, orchestrated teams of AI agents that can meet the demands of defending against AI-driven attacks by employing specialist tooling and skill sets.
CISOs also need to ensure that each agent has its own domain training and understanding, memory, and dedicated tools, and can act on its own within set limits, with a central coordinator that knows when to hand a decision to a human. This requires oversight, but in return, organizations will avoid agent sprawl, which the tech giant believes is the next iteration of shadow IT and its associated threats.
Please accept cookies to view the comments powered by Disqus.
Charlie Osborne is a cybersecurity journalist and photographer who writes for ZDNET and CNET from London. PGP Key: AF40821B See full bio
原文为英文,正文由自动抓取并保留原文。
原文发布于 ZDNet(5 ways AI will change how you defend your business in 2027),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。