本文来源:https://aimag.no/en/nyheter/gemini-agent-can-now-control-your-mac-google-warns-of-permanent-file-deletion
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Gemini 智能体现在可以操控你的 Mac——谷歌警告文件可能被永久删除
谷歌于 2026 年 11 月 19 日星期四在 macOS 版 Gemini 应用中推出了一项“计算机使用”(computer use)功能,让该助手能够在用户做其他事情时点击、输入并浏览本地文件、Mac 应用和网站。
谷歌 11 月 19 日宣布,macOS 版 Gemini 桌面应用现在可以代表用户完成多步骤任务。该智能体在一个独立的后台窗口中工作,点击菜单、表单和网页,而用户的屏幕和光标仍可自由使用。聊天窗口中会实时预览智能体正在执行的操作。TechCrunch(techcrunch.com)援引谷歌的公告报道了这一消息。
据 Let's Data Science 对谷歌博客文章的总结,用户描述一个多步骤任务,Gemini 便负责在本地文件、桌面应用和网页中进行点击、输入、操作菜单和填写表单。
该智能体在一个独立的后台窗口中运行,并在聊天窗口中提供实时预览。
TechCrunch 写道,用户可以随时暂停、停止智能体,并在操作执行前进行批准——与其他“计算机使用”智能体类似。
谷歌表示,“关键操作”需要用户事先批准。然而,公告并未定义什么算作关键操作——这是此次发布留下的待解问题之一。
据 TNW 报道,该智能体最初只能看到用户明确授权连接的文件夹。这限制了它在本地可浏览的范围,但并不提供完全的保护——请参阅下方的风险警告。
使用要求:Apple Silicon 芯片、macOS 15 及以上、付费订阅、英语
以下细节决定了大多数用户究竟能否使用这项功能:
一台搭载 Apple Silicon 芯片、运行 macOS Sequoia 15.0 或更新系统的 Mac。Intel 芯片的 Mac 不在支持之列。
一份付费的谷歌 AI 套餐:Plus、Pro 或 Ultra。该功能不免费提供。
各消息来源对可用地区说法不一。TechCrunch 称该功能“首先在美国以英语提供”。而 Let's Data Science 和 TNW 则描述了面向英语用户的更广泛推出,明确排除了欧洲经济区(EEA)、英国、瑞士和尼日利亚——据 TNW 称,与谷歌的 Gemini Spark 一样,该功能在欧洲和英国不可用。这一分歧尚未得到澄清;可以确定的是,挪威和其他欧洲用户目前只能等待,谷歌也尚未给出扩展日期。
据 TechCrunch 转述,所有示例均来自谷歌本身:填写表单、对一批文件重复执行同一任务、按特定标准将文件整理到文件夹中,以及根据文件或网页仪表盘中的信息制作幻灯片演示。换言之,这些都是繁琐、重复的工作任务——而非创造性或探索性的任务。
目前尚无独立测试衡量它的实际效果。谷歌的公告没有给出任何指标、成功率或完成时间,也没有提及该功能由哪款 Gemini 模型驱动(Let's Data Science)。到目前为止,读者唯一的参考依据就是该公司自己的描述。
公告中最引人注目的部分,是谷歌自己承认可能出的问题。据 TNW 报道,谷歌警告称 Gemini 可能会犯错误,包括永久删除文件或把它们分享给他人。该功能被设计为事先请求许可,但这一警告依然存在。
而且备份机制十分有限:Gemini 修改过文件的备份会在 24 小时后或用户开始新任务时被删除。换句话说:如果智能体错误地修改或删除了内容,用户只有极短的时间窗口来发现并撤销。这使得批准和暂停控制听起来没有那么正式——在实践中,它们才是最重要的安全网。
TechCrunch 指出,谷歌“入场稍晚”——OpenAI、Anthropic 和 Meta 早已在各自的桌面应用中提供智能体功能。谷歌在这项技术上其实耕耘已久:据 TNW 称,早在 2026 年 6 月,谷歌就在面向开发者的 Gemini 3.5 Flash 中内置了计算机使用功能。如今发生的是,一项成熟的功能正走向普通消费者——至少是其中付费、使用英语且拥有 Apple Silicon 芯片的那部分人。
此次发布与谷歌更广泛推出 Gemini Tasks 恰在同一周,这表明谷歌有意推动 Gemini 变得更具行动力——而不只是用于对话。
此次发布留下了若干未解问题,消息来源中均有记载:
哪款模型?谷歌并未说明是哪款 Gemini 模型驱动计算机使用功能。
效果如何?无论是谷歌还是独立测试,都没有给出指标、成功率或耗时。
“关键操作”是什么?批准机制确实存在,但定义缺失。
在哪里能用?TechCrunch 称首先在美国推出;另外两家来源则描述了除部分地区外的全球英语推出。目前缺乏明确澄清。
更多地区何时开放?没有日期,且据其中两家来源称,欧洲被明确排除在外。
使用额度?各套餐(Plus/Pro/Ultra)均未公布具体限制。
对于欧洲经济区以外符合条件的付费用户来说,此次发布意味着他们可以把繁琐的文件和表单工作交给智能体——有一个可见的安全网,但同时也伴随着谷歌自己关于永久数据丢失风险的警告。对于挪威和其他欧洲用户而言,目前则毫无影响:该功能被排除在外,且没有承诺任何开放日期。而从整个市场来看,此次发布表明谷歌在桌面环境的智能体功能方面仍是第三或第四名入场——尽管该公司早在 6 月就已掌握这项技术。
谷歌赋予 Gemini 替你使用 Mac 的能力 | TechCrunch——techcrunch.com
谷歌为 Mac 版 Gemini 应用添加计算机使用功能 | Let's Data Science——letsdatascience.com
谷歌的 Gemini 现在可以接管你 Mac 上的任务——thenextweb.com
本文来源:https://aimag.no/en/nyheter/gemini-agent-can-now-control-your-mac-google-warns-of-permanent-file-deletion
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Gemini agent can now control your Mac – Google warns of permanent file deletion
Google launched a "computer use" feature in the Gemini app for macOS on Thursday, November 19, 2026, allowing the assistant to click, type, and navigate local files, Mac apps, and websites while the user does something else.
Google announced on November 19 that the Gemini desktop app for macOS can now complete multi-step tasks on the user's behalf. The agent works in a separate background window, clicking through menus, forms, and web pages, while the user's screen and cursor remain free. A real-time preview of what the agent is doing appears in the chat window. TechCrunch reported the news ( techcrunch.com ), citing Google's announcement.
According to Let's Data Science , which summarizes Google's blog post, the user describes a multi-step task, and Gemini handles clicking, typing, menus, and forms across local files, desktop apps, and the web.
The agent works in a separate background window, with a live preview in the chat.
The user can pause at any time, stop the agent, and approve actions before they are carried out — similar to other "computer use" agents, TechCrunch writes.
Google says "key actions" require the user's advance approval. What counts as a key action, however, is not defined in the announcement — one of the open questions the launch leaves behind.
The agent can initially only see folders the user explicitly connects it to, TNW reports. That limits the scope of what it can navigate locally, but it does not provide full protection — see the risk warnings below.
The requirements: Apple Silicon, macOS 15+, paid subscription, English
Here are the details that determine whether most users can get the feature at all:
A Mac with Apple Silicon running macOS Sequoia 15.0 or newer . Intel Macs are excluded.
A paid Google AI plan : Plus, Pro, or Ultra. The feature is not available free of charge.
The geography is unclear in the sources. TechCrunch writes that the feature is "initially available in English in the US." Let's Data Science and TNW, on the other hand, describe a broader English-language rollout that explicitly excludes the EEA, the UK, Switzerland, and Nigeria — and which, according to TNW, like Google's Gemini Spark, is not available in Europe or the UK. The discrepancy has not been resolved; what is certain is that Norwegian and other European users must wait for now, and that Google has given no date for expansion.
The examples all come from Google itself, as relayed by TechCrunch: filling out forms, repeating the same task across a batch of files, organizing files into folders by specific criteria, and building slideshow presentations from files or information in web dashboards. In other words, tedious, repetitive work tasks — not creative or exploratory ones.
There are no independent measurements of how well this actually works. Google's announcement gives no metrics, success rates, or completion times, and does not mention which Gemini model powers the feature (Let's Data Science). The reader's only reference point so far is the company's own descriptions.
The most striking part of the announcement is what Google itself says can go wrong. According to TNW, Google warns that Gemini can make mistakes, including permanently deleting files or sharing them with others . The feature is designed to ask permission first, but the warning stands nonetheless.
And the backup scheme is narrow: backups of files Gemini has changed are deleted after 24 hours or when the user starts a new task. In other words: if the agent erroneously changes or deletes something, the user has a very short window to detect and reverse it. That makes the approval and pause controls less formal than they may sound — in practice, they are the most important safety net.
TechCrunch points out that Google is "arriving a little late" — OpenAI, Anthropic, and Meta have already offered agentic features in their desktop apps for some time. Google has been working on the technology for a long while: computer use was built into Gemini 3.5 Flash for developers back in June 2026, according to TNW. What is happening now is that an experienced feature is reaching ordinary consumers — or at least the paying, English-speaking, Apple Silicon-owning ones among them.
The launch also comes the same week as Google's broader rollout of Gemini Tasks, which suggests a deliberate push to make Gemini more action-oriented — not just conversational.
The launch leaves several unanswered questions, all documented in the sources:
Which model? Google does not say which Gemini model powers computer use.
How good is it? No metrics, success rates, or times — from neither Google nor independent tests.
What are "key actions"? The approval mechanism exists, but the definition is missing.
Where can you get it? TechCrunch says the US first; the other two sources describe a global English-language rollout with exceptions. Clarification is lacking.
When do more regions come? No date, and Europe is specifically excluded according to two of the sources.
Usage limits? No stated limits per plan (Plus/Pro/Ultra).
For eligible paying users outside the EEA, the launch means they can delegate tedious file and form work to an agent — with a visible safety net, but also with Google's own warning about permanent data loss in the background. For Norwegian and other European users, it means nothing for now: the feature is excluded, and no date has been promised. And for the market in general, the launch shows that Google is still third or fourth out of the gate with agentic features in the desktop environment — even though the company has had the technology since June.
Google gives Gemini the ability to use your Mac for you | TechCrunch — techcrunch.com
Google adds computer use to Gemini app for Mac | Let's Data Science — letsdatascience.com
Google’s Gemini can now take over tasks on your Mac — thenextweb.com
