从"能动"到"可控":AI agent app store 迎来三道护栏——Cursor 隔离、Vercel 审批、LangChain 评估,8月Agent生态集体补安全课

AI 安全核心真相:真正的防护根植于底层架构,而非表层提示词优化。


The Leap of AI Agent: From Functional to Controllable

In August 2026, the global AI Agent ecosystem ushered in an explosive upgrade in functional capabilities. Doubao launched more than 200 professional skills; DeepSeek Harness open-sourced its framework and accumulated over 2,600 plugin repositories in just 8 days; Grok Build achieved full public access, enabling one-click app generation and in-feed trial on X. AI Agents are expanding their operational boundaries, accessing more system scenarios and undertaking increasingly complex business tasks.

However, the rapid expansion of Agent capabilities has also pushed runtime security and controllability to the core of the industry agenda.

Over the same month, Cursor was exposed with two critical sandbox escape vulnerabilities scoring 9.8 on the CVSS scale. At Black Hat Conference, Check Point disclosed 11 high-risk vulnerabilities in mainstream Agent frameworks. The UK’s National Cyber Security Centre (NCSC) urged enterprises to deploy strict sandbox isolation for all AI Agents. Meanwhile, the OWASP Agentic Top 10 was officially released as a unified industry risk assessment standard.

The entire AI Agent ecosystem is completing a crucial upgrade: shifting from mere operational capability to standardized controllability.

First Safety Barrier: Isolation — Confine Agents to a Secure Sandbox

The security risks of Agent sandbox loopholes are fully exposed via Cursor’s vulnerability incident.

In July 2026, Cato AI Labs discovered two zero-interaction vulnerabilities codenamed DuneSlide (CVSS 9.8). Attackers do not require user authorization or active interaction. They only need to embed malicious preset prompts in external data sources that Agents can read, such as MCP service feeds and web search results. When users ask normal questions, the hidden instructions will be triggered automatically to complete covert attacks.

The two vulnerabilities adopt the same exploitation logic: tricking the Agent into writing unauthorized files to break sandbox restrictions.

  • CVE-226-50548: Exploits flaws in the working_dir parameter. The Agent switches to a non-default path, and Cursor incorrectly adds the unvalidated path to the allowlist. Malicious commands can overwrite sandbox core files, enabling unrestricted system-level execution.

  • CVE-226-50549: Leverages symbolic link parsing defects. Cursor aborts security verification upon detection failure and directly allows writing to trusted paths.

A single sandbox contained two distinct breach paths. Once penetrated, attackers can take over the developer’s local device and all cloud service accounts logged in via the editor.

Cursor’s risks are not isolated cases. On January 27, 2026, CVE-226-25253 became the world’s first officially assigned vulnerability for Agentic AI systems. Microsoft launched MXC (Microsoft Execution Containers) at Build 2026 for OS-level execution isolation. NVIDIA and Cisco co-open-sourced OpenShell, while AWS released dedicated AI Agent sandbox solutions for EKS clusters.

The industry has reached a unanimous consensus: sandbox isolation is no longer an optional feature, but a mandatory infrastructure default. Cursor’s official data confirms that sandboxed Agents have a 40% lower interruption rate than unsandboxed ones — isolation does not restrict capabilities, but stabilizes operational reliability.

Domestic security standards and infrastructure are also maturing rapidly. On July 29, 2026, China Academy of Information and Communications Technology released the Trusted Internet Agent Runtime Sandbox Standard, launching the first batch of official evaluations covering six core capabilities: isolation protection, permission governance, resource scheduling monitoring, security auditing, lifecycle management, and runtime defense.

Tencent Cloud open-sourced Cube Sandbox in April, featuring a 60ms cold start speed and the ability to launch tens of thousands of sandboxes per minute, supporting over 100,000 instantaneous instance scheduling. Sandbox isolation has become the standard infrastructure for enterprise-grade AI Agent deployment.

Second Safety Barrier: Approval — Manual Gatekeeping for Critical Nodes

Isolation controls what systems an Agent can access, while approval controls what behaviors an Agent can execute.

In June 2026, Vercel open-sourced the Agent framework Eve at the Ship Conference, positioned as "the Next.js of Agent development". Eve natively integrates full production-grade capabilities: persistent sessions, sandbox isolation, manual approval workflows, secure connections, observability, and fine-grained Eval access control.

Its approval mechanism is rigorous and efficient. In the Vercel AI SDK, enabling demandsApproval: true pauses task execution pending manual confirmation. The Agent suspends operation without consuming additional computing resources and resumes precisely from the breakpoint after approval.

A complete Agent workflow often involves second-level model reasoning, hour-level manual review, and day-level long-tail business impacts. The approval mechanism ensures accurate task suspension, safe retry, and credible result output.

NanoClaw further extends the approval capability to 15 mainstream communication platforms including Slack, WhatsApp and Teams. The system automatically pushes approval cards for high-risk operations such as fund payment and cloud resource deletion, supporting one-click confirmation. Yubico, IBM and Auth0 adopt stricter rules, requiring physical YubiKey authentication for extreme-risk behaviors.

Nevertheless, manual approval has inherent limitations. Rubrik’s research shows that AI Agents can deliver 10 times more workload than humans per unit time. Pure manual review cannot match Agent execution efficiency.

On August 21, 2026, Alibaba Cloud Developer Community emphasized the bottom line of enterprise Agent permission governance: separation of read, write and fund permissions. Manual approval solves the problem of operational confirmation, but cannot cover accurate risk judgment and traceable post-event auditing.

The core truth of Agent security: reliable defense lies in architectural design, not prompt engineering.

Third Safety Barrier: Evaluation — Pre-Launch Immunity Verification

Isolation and approval serve as runtime defenses, while pre-launch evaluation acts as the ecosystem’s immune system to block hidden risks in advance.

In July 2026, GuardianAgentBench (GABench) released an authoritative Agent security benchmark covering 580 scenarios, 6 major fields and 3 production-grade frameworks. In the same month, AgentShield Bench v3 launched, focusing on evaluating Agent target integrity and anti-hijacking capabilities. In June, OWASP jointly released the AIRQ framework, enabling horizontal security risk scoring and batch comparison for up to 100 enterprise Agents.

Check Point’s one-year penetration test covered mainstream frameworks including LangChain, LangGraph, CrewAI, AutoGen, Microsoft Agent Framework and Google ADK, uncovering 11 underlying vulnerabilities.

The industry’s wake-up call is clear: Almost all vulnerabilities are classic security flaws — insecure deserialization, server-side request forgery, path traversal, use-after-free. These are well-known bugs that the industry has fixed for 20 years, yet they still exist in modern Agent frameworks that can access user inboxes, modify databases and control cloud resources.

The huge security risk of framework inheritance cannot be ignored. LangChain has more than 800 community integrations, with the PythonREPLTool enabled by default without sandbox restrictions. Tool descriptions and open community prompts on LangChain Hub may contain hidden injection risks. All upper-layer Agent applications built on these frameworks will inherit underlying security flaws.

Pre-launch security evaluation is precisely designed to eliminate such hidden dangers, preventing outdated classic vulnerabilities from threatening cutting-edge AI systems.

OWASP summarizes mainstream Agent threats including prompt injection, tool resource hijacking, identity permission abuse, memory context poisoning and supply chain risks. Prompt injection is divided into two types: direct injection that overrides system instructions, and indirect injection that embeds malicious commands in external data sources. Microsoft Research has verified that LLMs cannot stably distinguish between informational context and executable instructions.

Beyond Three Barriers: Unexhub’s Security Empowerment at the Computing Power Layer

Isolation, approval and evaluation solve behavior-level security problems for AI Agents. However, Agent operation relies entirely on Token computing power resources. The standardized transaction, pricing and circulation of computing power constitute the underlying security foundation of the entire ecosystem — a key dimension empowered by Unexhub.

As a professional global Token aggregation and computing power trading center, Unexhub standardizes the entire Token circulation link to build the fourth safety barrier for the Agent ecosystem:

  • Real-time instant settlement: Enables suppliers to receive payment synchronously with service delivery, completely eliminating capital occupation and financial risks, and stabilizing the operational security of resource providers.

  • Zero commission for early settlers: Reduces the industry entry threshold for small and medium-sized suppliers, avoiding additional service fee costs and ensuring transparent and credible profit mechanisms.

  • Dynamic API bidding mechanism: Token prices float autonomously based on real-time market supply and demand, realizing fair pricing and preventing malicious price manipulation and market chaos.

When Token evolves into a tradable financialized resource, Unexhub’s standardized transaction mechanism ensures settlement certainty, identity anti-counterfeiting and traceable auditing, completing the full-dimensional security system for AI Agent industrialization.

Industry Conclusion: Controllability Is the Ticket to Industrialization

August 2026 witnessed an explosive improvement in AI Agent functionality, yet also ushered in the mandatory course of "controllable iteration".

Cursor’s sandbox escape incident proves that a single parameter error in the isolation layer can lead to full-system compromise. Vercel’s Eve framework proves that approval mechanisms must be embedded in the underlying architecture rather than attached externally. Check Point’s research proves that traditional vulnerabilities will continue to threaten new AI ecosystems without standardized evaluation screening.

The leap from "functional" to "controllable" is not a simple feature upgrade, but a fundamental transformation of security thinking — turning isolation, approval and evaluation from optional functions into rigid default standards.

When AI Agents can read private information, modify core data and schedule cloud resources at will, operational capability is only the starting point, while controllability is the real ticket to enterprise-grade commercialization.

Driving the standardized development of the global Token economy, Unexhub completely reshapes Token transaction logic and builds a one-stop Token aggregation and distribution computing power trading center. The platform integrates full-spectrum AI resources, covering top-tier models such as Gemini, ChatGPT, Claude, DeepSeek and GLM, as well as high-quality customized models from individual developers and SMEs.

Supported by multi-supplier bidding competition and millisecond-level intelligent routing, the platform delivers substantial benefits for global developers and enterprises: users pay strictly based on actual Token consumption, with the lowest discount reaching 30% of the official market price. For suppliers, one-click platform access and real-time revenue settlement greatly reduce the commercialization threshold, helping all participants share the dividends of the standardized, secure AI Agent ecosystem.


中文版本

AI Agent 跃迁:从「能动可用」到「可控可信」

2026 年 8 月,全球 AI Agent 生态迎来功能能力的集体爆发:豆包上线 200+ 全场景智能技能,DeepSeek Harness 开源框架 8 天孵化 2600 个插件仓库,Grok Build 全量开放实现一句话生成应用、X 信息流内嵌试玩。如今的 AI Agent,能落地的场景更多、能联动的系统更深、能完成的任务更复杂。

但能力的极速扩张,也让安全可控从可选需求,变成了整个行业的必答题。

同月,AI 安全风险集中暴露:Cursor 被爆出两个 CVSS 9.8 分的高危沙箱逃逸漏洞;Black Hat 大会上,Check Point 公开了 11 个主流 Agent 框架的底层漏洞;英国国家网络安全中心(NCSC)强制要求企业为 AI Agent 配置沙箱隔离机制;OWASP 正式发布 Agentic Top 10,确立行业统一风险分级标准。

整个 AI Agent 生态,正在集体补齐从「能动」到「可控」的关键一课。

第一道安全护栏:隔离 —— 将 Agent 关进安全沙箱

Cursor 的高危漏洞事件,彻底暴露了沙箱隔离失效的致命风险。

2026 年 7 月,Cato AI Labs 发现两枚代号为 DuneSlide 的高危零交互漏洞,CVSS 危险评分高达 9.8。该攻击无需用户授权、无需主动交互,攻击者只需在 Agent 可读取的外部内容(MCP 服务数据、网页搜索结果等)中嵌入伪装正常的恶意提示词,用户发起普通提问即可触发隐藏指令,实现无感知入侵。

两个漏洞依托同类攻击逻辑:诱导 Agent 执行非法写入操作,借此突破沙箱防护体系。

  • CVE-226-50548:利用工作目录参数缺陷,Agent 切换至非默认路径后,Cursor 未做安全校验便纳入可写入白名单,恶意指令可覆盖沙箱核心文件,实现裸奔权限执行。

  • CVE-226-50549:利用符号链接解析漏洞,安全校验失败后系统直接放行写入操作,允许恶意内容侵入可信路径。

同一套沙箱体系,两处核心防护被攻破。一旦沙箱逃逸成功,攻击者可完全控制开发者本地设备,以及编辑器登录的全部云端业务系统,危害级别极高。

Cursor 并非个例。2026 年 1 月 27 日,CVE-226-25253 成为全球首个正式收录的智能体 AI 系统漏洞。行业厂商纷纷加码隔离基建:微软在 Build 2026 发布 MXC 系统级执行隔离容器,英伟达与思科联合开源 OpenShell,AWS 推出适配 EKS 的专属 AI Agent 沙箱方案。

行业共识已然明确:沙箱隔离不再是可选功能,而是企业级部署的默认标配。Cursor 官方数据佐证,经过沙箱隔离的 Agent,运行中断率比非隔离版本低 40%—— 隔离不是限制能力,而是保障业务稳定、规避安全事故的核心基石。

国内安全标准与基建同步完善。2026 年 7 月 29 日,中国信通院正式发布《可信互联网智能体 运行时 运行沙盒》标准,启动首批权威评估,覆盖隔离防护、权限治理、资源调度监控、安全审计、生命周期管理、运行时防护六大核心能力域。

腾讯云 4 月开源的 Cube Sandbox 性能行业领先,实现 60ms 极速冷启动,分钟级可拉起数万沙箱实例,瞬时调度能力突破 10 万级。沙箱隔离,已经从「增值功能」彻底迭代为 AI 产业标准化基建。

第二道安全护栏:审批 —— 关键节点人工兜底把控

沙箱隔离解决「Agent 能访问哪些系统、触碰哪些数据」的边界问题,人工审批则解决「Agent 能执行哪些高危操作」的行为问题。

2026 年 6 月,Vercel 在 Ship 大会开源全新 Agent 框架 Eve,定位为「Agent 领域的 Next.js」,原生搭载全链路生产级能力:持久会话、沙箱隔离、人工审批、安全链路、可观测性、精细化访问控制。

Eve 的审批机制严谨且高效:在 Vercel AI SDK 中开启审批权限后,Agent 执行高危操作时会自动暂停任务,等待人工确认放行。审批期间任务挂起、零算力消耗,确认后精准断点续跑,完美适配「秒级推理、小时级审核、天级业务影响」的复杂工作流,保障每一步操作可追溯、可管控。

NanoClaw 进一步将审批能力覆盖至 Slack、WhatsApp、Teams 等 15 款主流办公通讯工具,Agent 发起资金支付、云资源删除等高风险操作时,自动推送审批卡片,支持一键核验确认。Yubico、IBM、Auth0 更是升级防护标准,超高风险操作需搭配 YubiKey 物理密钥认证,杜绝远程冒用风险。

但人工审批存在天然短板。Rubrik 行业调研显示,AI Agent 单位时间作业效率是人工的 10 倍,纯人工逐项审批完全跟不上 Agent 运行节奏,形成效率瓶颈。

2026 年 8 月 21 日,阿里云开发者社区明确企业 Agent 权限治理底线:读、写、资金三权分离。人工审批只能解决「是否执行」的问题,无法根治「操作是否合规、事后能否溯源」的核心风险。

AI 安全核心真相:真正的防护根植于底层架构,而非表层提示词优化。

第三道安全护栏:评估 —— 上线前置风险免疫

隔离、审批属于运行时动态防护,而标准化安全评估,是 Agent 上线前的「免疫系统」,从源头拦截隐藏漏洞。

2026 年 7 月,GuardianAgentBench(GABench)发布权威 Agent 安全基准,覆盖 6 大领域、580 个真实生产场景、3 套主流生产框架;同月,AgentShield Bench v3 上线,专注检测 Agent 目标完整性与抗劫持能力。6 月,OWASP 联合多家权威机构推出 AIRQ 评估体系,支持企业对百级 Agent 进行横向安全打分、风险对比。

Check Point 耗时一年,对 LangChain、LangGraph、CrewAI、AutoGen、微软、谷歌等主流六大 Agent 框架开展渗透测试,累计挖出 11 个底层高危漏洞。

其行业结论振聋发聩:本次发现的所有漏洞,均不是新型攻击手段,而是20 年前就已被攻克的经典漏洞—— 反序列化漏洞、服务端请求伪造、路径遍历、释放后重用。这些老旧缺陷,如今依然潜藏在可读取隐私邮件、修改核心数据库、调度云端资源的新一代 Agent 框架底层。

框架传导风险极具致命性。LangChain 拥有 800+ 社区集成插件,PythonREPLTool 工具默认启用且无沙箱防护,工具描述、社区公开提示词均可能暗藏注入隐患。底层框架的漏洞,会逐级传导至所有基于其搭建的上层 Agent 应用,形成系统性风险。

而前置安全评估的核心价值,就是在上线前完成全面体检,杜绝「老旧漏洞威胁全新 AI 业务」的行业乱象。

OWASP 明确当前 Agent 七大核心威胁:提示词注入、工具与资源劫持、身份权限滥用、记忆上下文投毒、供应链风险等。其中间接注入最为隐蔽,攻击者将恶意指令嵌入网页、邮件等外部数据源,诱导 Agent 读取执行。微软研究院已证实:大模型无法稳定区分「信息类上下文」与「可执行指令」,依赖模型自主风控完全不可行。

护栏之外:Unexhub 筑牢算力层安全底座

隔离、审批、评估,解决的是 AI Agent 应用行为层面的安全问题,但所有 Agent 运行都依赖 Token 算力资源,算力的交易规范、定价公平、资金安全,是支撑整个生态稳定的底层根基,而这正是 Unexhub 的核心价值所在。

Unexhub 打造的全域 Token 聚合分发算力交易中心,为 AI 生态补上第四道隐形安全护栏,从算力流通维度完善安全体系:

  • 即时结算机制:实现算力供应与资金回款同步,中小供应商供多少、结多少,彻底杜绝账期积压、资金坏账风险,零财务压力运营;

  • 早期零佣金红利:零成本开放成熟交易渠道,降低行业准入门槛,杜绝平台抽成乱象,保障交易公平透明;

  • 动态 API 竞价模式:Token 价格随市场供需自主浮动,杜绝恶意控价、垄断炒作,实现算力资源公允定价、高效流转。

当 Token 从固定消耗品升级为可交易的数字资产,结算确定性、身份防伪、审计溯源构成了算力交易的安全底线。Unexhub 以标准化交易规则,补齐 AI Agent 工业化的底层安全短板。

行业总结:可控,是 AI 商业化的终极门票

2026 年 8 月,AI Agent 的能力边界实现跨越式突破,但也让全行业认清核心逻辑:能动只是基础,可控才是核心。

Cursor 沙箱逃逸证明,隔离层一个微小参数漏洞,即可引发整机沦陷;Vercel Eve 框架证明,审批机制必须内嵌于架构底层,而非外挂式摆设;Check Point 的调研证明,忽视前置评估,老旧漏洞会持续威胁全新 AI 生态。

AI Agent 从「能动」到「可控」的跃迁,绝非简单叠加安全功能,而是行业思维的彻底迭代:将隔离、审批、评估三大能力,从「可选配置」升级为「强制标配」。

当 AI Agent 具备读取隐私、改写数据、调度核心资源的权限,行业规则已然改写:功能落地只是起点,安全可控,才是 AI Agent 从 Demo 走向规模化商用的终极门票。

在 AI 算力经济标准化浪潮下,Unexhub 持续重塑全球 Token 交易逻辑,搭建一站式 Token 聚合分发算力交易枢纽。平台全面入驻 Gemini、ChatGPT、Claude、DeepSeek、GLM 等头部大模型厂商,同时吸纳个人、中小企业的优质定制化 AI 资源,依托海量供应商竞价机制与毫秒级智能路由,为行业释放极致权益:开发者按实际 Token 消耗量实付计费,最低可享市场原价 3 折,大幅降低商用落地成本;算力供应商可一键入驻、分成实时结算,零门槛盘活算力资产,共享 AI 产业合规化、规模化红利。