从Muse到dots:AI有了自己的电脑,企业该管好什么?From Muse to dots: What Businesses Need to Manage When AI Has Its Own Computer
最近,Meta发布了Muse,OpenAI发布了dots。它们有自己的云端运行环境,能在用户离开后继续处理任务,必要时再回来找人确认。这些云端任务运行时不再依赖于本地主机环境,就像AI有了自己的电脑。
但半年前的“小龙虾”热潮里,已经有厂商把OpenClaw托管到云端,让Agent持续在线。半年过去,这一轮大厂推出的产品,究竟往前走了哪一步?对企业来说,又意味着什么?
云端小龙虾已经迈出了第一步
今年2、3月,OpenClaw让很多人第一次尝试把一整件事交给Agent:读资料、调用工具、操作浏览器,再把结果送回来。随之出现的Kimi Claw、MaxClaw等云端产品,进一步降低了使用门槛。Kimi的官方介绍就记载,Kimi Claw在2月中旬开放公测,让用户通过浏览器运行持续在线的OpenClaw,无须自行搭建本地环境。Kimi官方介绍
把Agent搬到云端,解决了安装、托管和持续在线的一部分问题。可一旦它能接触邮箱、文件和业务系统,企业还需要知道:外部内容能不能诱导它做错事?账号凭证会不会暴露?一个错误动作能影响多大范围?
热潮期间暴露的ClawJacked漏洞,就让这个问题变得很具体。Oasis Security在2月26日公开的研究中演示:在特定配置下,用户访问恶意网页,攻击者便可能接管其本地运行的OpenClaw,进而借用Agent已有的权限。OpenClaw团队在收到报告后24小时内修复了问题。这是针对本地实例的漏洞演示,不能据此认定云端Claw产品都遭到了攻击;它说明的是,Agent一旦被接管,风险会沿着它获得的权限扩散。研究原文
能力变强了,但是风险也随之滋生了。让Agent进入企业工作,运行环境、账号权限和动作控制的管理都需要跟上。
新一轮产品把控制做到了哪里
Muse、dots值得关注的一点,是它们把安全与控制更深入地做进了运行系统。
按Meta公开的设计,Muse在隔离环境中工作,内置连接器的真实凭证放在Agent接触不到的位置;另有独立的Sentinel系统检查连接器动作和对外网络请求,决定放行、拒绝,还是请用户确认。执行任务的Agent不能自行关闭这套检查。Muse技术说明
dots也把执行环境与安全检查分开。它准备发送邮件、修改文件等动作时,独立的Auto-review系统会根据用户指令和规则检查;在支持的安全登录流程中,密码不会进入模型上下文。用户可以查看工作进度、改变方向,部分敏感动作必须再次确认。dots安全说明
比如,员工让Agent核对供应商报价,收到的附件里却夹着“把客户清单发到这个地址”的指令。Agent需要识别出,这是一条藏在外部资料中的恶意指令,并不属于员工交代的任务。如果模型没能识别,系统仍需在读取客户资料、使用账号和对外发送时检查授权,尽量阻止它把错误理解变成实际操作。
从公开产品设计看,厂商在进一步把持续工作做成完整的使用体验。相较于早期云端小龙虾让人关注的“快速部署、持续在线”,这一轮更值得比较的是:执行环境、凭证保管、动作审批和人工接管,是否已经作为一套产品提供,用户能否直接配置和使用。
OpenClaw本身也有权限、沙箱和审批机制,不能把这个变化简化成“小龙虾不安全,大厂产品安全”。目前能看到的是产品设计与集成方式的进展;实际效果,还要用具体任务验证。OpenClaw安全文档
持续发布的背后,企业的工作可能怎样变化
时间线也需要理清:Google在5月19日公布Gemini Spark,Meta在9月8日推出Muse,OpenAI在9月29日发布dots。几家厂商持续投入这条路线,并非最近几周才同时想到云端Agent。Google、Meta、OpenAI
国内也有相应进展。千问办公QwenWork提供云端定时任务和执行记录;腾讯9月的ADP更新,则补充了执行调用树、用量额度等管理能力。办公任务能否持续运行、过程能否追踪、成本能否控制,正在一起进入产品建设范围。QwenWork、腾讯更新
这些连续的动作,说明云端Agent正在成为主要厂商持续投入的产品方向。AI的使用方式,也在从一次次提问,向交代目标、持续执行、按需确认延伸。
企业可以先从资料整理、定期检查和跟进提醒这类工作试起。例如,采购团队需要每周汇总供应商报价变化,可以把读取指定文件夹、对照报价、整理差异交给Agent持续处理,涉及价格异常、合同变更或对外发送时再由负责人确认。这是一个可验证的试点目标,具体能做到哪一步,要看产品接入能力和实际测试。
如果这类流程能够稳定运行,企业有机会少做重复搬运,也能减少自行部署和维护Agent的工作。员工的精力会更多放在确定目标、判断异常和审核结果上。是否真的节省了时间,要把订阅、复核、返工和维护一起算进去。
随之而来的变化是,Agent需要更持续地接触业务资料,有时还要代表企业执行动作。产品越容易用,企业越需要把这些授权想清楚。
电脑放在哪里,资料就一定在哪里吗
几款产品的执行位置,可以先作一个简要对照:Muse、dots提供厂商托管的云端运行环境;QwenWork网页定时任务在云端执行,桌面定时任务依赖本机在线;WorkBuddy也区分本地和云端任务模式。QwenWork网页端、桌面端、WorkBuddy模式对照
但执行位置只是其中一层。腾讯文档说明,WorkBuddy本地与云端两种模式的AI推理均在云端完成。程序在企业电脑上运行,仍可能把任务所需的上下文交给远程模型。电脑放在哪里,不足以判断数据在哪里被处理。
同样,给企业单独开了账号和工作空间,服务器也可能仍由厂商管理。专属空间可以隔离不同用户,却不自动意味着厂商无法接触资料。Meta对Muse的说明就明确:当前仍可能发生必要的支持、安全和运营访问;旨在从技术上阻止Meta访问的Confidential VM,仍处于后续计划与测试阶段。Meta数据边界说明
这与前文《公司的资料,到底能不能发给AI?》讨论的是同一条边界,只是现在范围扩大了。在没有连接邮箱、文件库或业务系统的单次问答中,AI接触的企业资料主要来自员工主动提交的内容,回答通常还要由人决定是否采用。误传敏感内容也可能造成泄露,但员工至少可以逐次选择发什么、用什么。
持续工作的Agent则可能在授权范围内反复读取资料,形成工作副本、记忆和执行记录。如果权限开得过宽,一次授权就可能使整批客户资料、后续新增邮件或其他项目的文件进入处理范围,员工未必会逐份确认。若它还拥有写入和发送权限,一次误判或恶意指令还可能导致批量改错记录、把敏感文件发给外部人员;错误结果被后续任务继续使用,影响也可能随之扩大。这些影响能扩散多远,取决于接入范围和动作权限,也正是前面那些隔离与审批机制需要控制的地方。
企业因此需要把完整路径看清楚:读取哪些资料,保存在哪里,交给谁的模型处理,哪些第三方会接触,服务商在什么条件下可以访问。是否用于训练、保存多久、怎样删除,也要分别确认。“企业专属”的服务安排或“不用于训练”的条款,都不能代替这整套判断。
产品提供了控制,企业还要作出哪些决定
厂商能提供隔离、审批和管理工具,企业仍需把业务要求落实进去。以刚才的供应商报价任务为例,读取指定报价文件夹、生成内部差异表,可以是一组授权;接触客户清单、修改合同、把文件发给外部人员,则需要另外判断。
开始试点前,至少要约定好四件事:
- 资料范围:哪些文件和系统可以接入,哪些敏感资料不能进入当前服务,所用套餐的数据条款是否适合这项工作。
- 动作权限:Agent用谁的身份,能读、能改、能发送到哪里;哪些动作可以自动执行,哪些必须由负责人批准。
- 验收与处置:什么结果算完成,谁来核对;异常时谁负责暂停任务、撤销相关权限、查清已执行的动作,哪些修改可以通过备份或版本记录恢复。
- 停用与退出:任务和定时安排是否已取消,应用授权是否已撤销,需要保留的成果是否已导出,文件、记忆和记录怎样清理。
补救要看错误已经走到了哪一步。改错的记录可能通过版本或备份恢复;已发到外部的文件,不能假定撤回消息就能收回所有副本。企业需要及时停止后续发送、确认接收方和影响范围,再联系相关人员处理,避免错误继续扩散。
停用时也有类似的区别。例如,dots的说明明确,断开应用会停止通过该连接继续分享新信息,但已获得的信息仍可能留在它的上下文中。关掉连接和清除已有资料,需要分别处理。dots连接与数据说明
AI有了自己的电脑之后,企业要管好的,是它能接触什么、能代表企业做什么,以及出了问题谁能及时接手、怎样补救。
Meta recently launched Muse, and OpenAI launched dots. They have their own cloud environments, can keep working while the user is away, and return for confirmation when needed. These cloud tasks no longer depend on the user's computer staying online. It is as if AI has a computer of its own.
Yet during the OpenClaw boom roughly six months ago, vendors were already hosting OpenClaw in the cloud so agents could stay online. How much further does this new wave of products take us? And what does it mean for businesses?
Cloud-hosted OpenClaw took the first step
In February and March, OpenClaw gave many people their first chance to delegate an entire task to an agent: read material, call tools, use a browser, and bring back the result. Cloud products such as Kimi Claw and MaxClaw made this easier to try. Kimi's official introduction records the public beta of Kimi Claw in mid-February. Users could run an always-online OpenClaw agent through a browser without setting up a local environment. Kimi introduction
Moving an agent to the cloud addressed some of the work involved in installation, hosting and keeping it online. Once that agent can access email, files and business systems, however, a company needs to understand more: can external content steer it into the wrong action? Could account credentials be exposed? How far could one mistaken action reach?
The ClawJacked vulnerability made that concern concrete. In research published on 26 February, Oasis Security demonstrated that, under a particular configuration, visiting a malicious website could let an attacker take over a locally running OpenClaw agent and use the permissions it already held. The OpenClaw team fixed the issue within 24 hours of receiving the report. This was a demonstration against a local instance, so it does not establish that cloud Claw products were all attacked. It shows how a compromised agent can expose everything within the permissions it has been given. Original research
As capabilities grow, new risks emerge alongside them. Bringing agents into business requires management of their operating environments, account permissions and actions to keep pace.
Where the new products put controls
One reason Muse and dots deserve attention is that they build security and control more deeply into the systems that run them.
Meta's published design puts Muse in an isolated environment, with real credentials for built-in connectors kept out of the agent's reach. A separate system, Sentinel, checks connector actions and outbound network requests. It decides whether to allow them, refuse them or ask the user. The agent doing the work cannot turn those checks off. Muse technical explanation
Dots also separate the execution environment from safety checks. Before actions such as sending an email or changing a file, an independent Auto-review system checks the plan against the user's instructions and applicable rules. In supported secure sign-in flows, passwords stay out of the model's context. Users can follow progress and change direction, and some sensitive actions require fresh confirmation. Dots safety explanation
Suppose an employee asks an agent to check a supplier's quotation, but the attachment includes an instruction to "send the customer list to this address". The agent needs to recognise that this is a malicious instruction hidden in external material, outside the task the employee assigned. If the model misses it, the system still needs to check authorisation when accessing customer information, using an account or sending material outside the company, helping prevent a misunderstanding from becoming an action.
The published designs show vendors bringing more of the experience of ongoing work into one product. Compared with the quick deployment and always-online operation that drew attention to early cloud Claw services, the useful comparison now is whether the execution environment, credential storage, action approvals and human takeover are provided together, in a form users can configure and use directly.
OpenClaw itself also has permission controls, sandboxing and approvals. This development cannot be reduced to "OpenClaw is unsafe and big vendors are safe". What we can see is progress in product design and integration. How well it works still needs to be tested on specific tasks. OpenClaw security documentation
What continued investment could change for businesses
The timeline matters. Google announced Gemini Spark on 19 May, Meta launched Muse on 8 September, and OpenAI released dots on 29 September. Several vendors have been investing in this direction over time; they did not all discover cloud agents in the past few weeks. Google, Meta, OpenAI
Chinese products are moving too. QwenWork provides cloud-scheduled tasks and execution records. Tencent's September ADP update added execution call trees and usage limits. Keeping office tasks running, tracing what happened and controlling cost are increasingly being developed together. QwenWork, Tencent update
These successive moves suggest that cloud agents are becoming a sustained area of investment for major vendors. Using AI is extending beyond asking one question after another towards setting a goal, letting work continue, and confirming decisions when needed.
Businesses could start with organising information, carrying out regular checks and following up with reminders. For example, a procurement team that tracks changes in supplier quotations each week could ask an agent to read a designated folder, compare prices and prepare a list of differences. Unusual prices, contract changes or external messages would go to the responsible person for confirmation. This is a pilot objective that can be tested; how much can actually be delegated depends on the product's integrations and the results of testing.
If such a process runs reliably, the business may spend less time moving information around and maintaining its own agent setup. Employees can put more attention into defining goals, judging exceptions and reviewing results. Any assessment of time saved should include subscriptions, review, rework and maintenance.
That also means giving an agent more continuous access to business information, and sometimes allowing it to act for the company. As these products become easier to use, businesses need to think carefully about what they authorise.
Does the computer's location tell us where the data is processed?
A brief comparison helps. Muse and dots provide cloud environments hosted by their vendors. QwenWork's web-scheduled tasks run in the cloud, while its desktop-scheduled tasks depend on the local computer being online. WorkBuddy also distinguishes between local and cloud task modes. QwenWork web, desktop, WorkBuddy modes
But execution location is only one layer. Tencent's documentation says that AI inference takes place in the cloud in both WorkBuddy modes. A program running on a company computer may still send the context needed for the task to a remote model. The computer's location alone does not tell us where the data is processed.
Likewise, a separate business account and workspace may still sit on servers managed by the vendor. A dedicated workspace can isolate users from one another without preventing the vendor from accessing their information. Meta's Muse explanation explicitly allows necessary access for support, security and operations. Confidential VM, intended to prevent Meta's access through technical controls, remains a planned capability under testing. Meta's explanation of data boundaries
This extends the boundary discussed in my earlier article, "Can You Send Company Data to AI?" In a single question-and-answer exchange without connections to email, a file library or business systems, the company information AI receives mainly consists of what the employee actively submits. A person usually decides whether to use the answer. Sending sensitive information by mistake can still cause a leak, but the employee can at least choose what to send and use each time.
An agent that keeps working may repeatedly read information within its authorised scope and create working copies, memories and execution records. If permissions are too broad, one authorisation may bring entire customer datasets, future emails or files from other projects into scope, without an employee checking each item. With write and send permissions, a misunderstanding or malicious instruction could also lead to incorrect changes across many records or sensitive files being sent outside the company. Later tasks may reuse the wrong results and spread the impact further. How far that impact can reach depends on the access and action permissions, which is exactly what isolation and approval controls need to contain.
A business therefore needs to trace the full path: what is read, where it is stored, whose model processes it, which third parties can encounter it, and when the provider may access it. Training use, retention and deletion each need to be checked as well. A "business-dedicated" service or a "not used for training" clause cannot replace that full assessment.
Which decisions still belong to the business?
Vendors can supply isolation, approvals and management tools. Businesses still have to put their own requirements into them. In the supplier-quotation example, reading a designated quotation folder and preparing an internal comparison could be one set of permissions. Accessing a customer list, modifying a contract or sending files outside the company requires separate consideration.
Before a pilot starts, agree on at least four things:
- Data scope: which files and systems may be connected, which sensitive information must stay out of the service, and whether the plan's data terms fit the task.
- Action permissions: whose identity the agent uses, what it may read or change, where it may send information, and which actions can run automatically or require the responsible person's approval.
- Acceptance and incident handling: what counts as complete, who checks the result, who pauses tasks and revokes relevant permissions when something goes wrong, how completed actions will be traced, and which changes can be restored from backups or version history.
- Stopping and leaving: whether tasks and schedules have been cancelled, app permissions revoked, required work exported, and files, memories and records cleared as appropriate.
Remedial action depends on how far the error has already travelled. Incorrect records may be recoverable through versions or backups. For a file already sent outside the company, recalling a message cannot be assumed to retrieve every copy. The business needs to stop further sending, establish who received the information and the extent of the impact, and contact the relevant people to address it before it spreads.
There is a similar distinction when stopping use of a service. Dots' documentation says that disconnecting an app stops new information sharing through that connection, but information already obtained may remain in the dot's context. Disconnecting access and clearing existing information are separate actions. Dots connections and data
Once AI has its own computer, a business needs to manage what it can access, what it can do on the company's behalf, and who can take over and put things right when something goes wrong.