product

2026-11-17 18:00

Early this morning, Apple pushed macOS 27.3 Beta 2 to developers. After upgrading, many developers found that when the application applied for "full disk access permissions", the system would pop up a warning window with a red exclamation mark:

The pop-up window pretentiously lists the potentially exposed emails, messages, photos and browsing history one by one, and requires you to confirm the risks one by one before you can finally click "Allow" to let it go.

Just like a full-screen warning that pops up when sensitive permissions are enabled on a domestic Android phone. However, Apple has not yet used the 10s countdown to “hard control” users.

In addition to warnings, macOS 27.3 Beta 2 also adjusts the management of permissions such as "Device Control and Data Access".

Users can be more precise about which folders or apps they allow access to, or they can ask the app to ask again each time they access it. Once an app controls other apps in the background, a corresponding prompt will appear in the upper right corner.

Swipe up and down to view more content

What prompts Apple to do this is precisely the increasingly capable AI Agent.

A month ago, Apple issued a brief developer announcement stating that some developers were using full disk access in ways that could compromise user data, and clearly mentioned the risks posed by the increased autonomy of AI Agents.

In just over a month, macOS 27.3 Beta 2 has implemented these changes into the system.

Revisit Mac permissions

Mobile phones and computers originally have two completely different permission logics.

Apple pays more attention to the separation of each app on the iPhone and iPad. Each app is restricted to its own "small room", that is, the "sandbox". Under normal circumstances, no App can freely access the data of other Apps.

But the Mac was a more open-purpose computer from the beginning.

Users can install apps from outside the App Store, manipulate the file system directly, execute commands using the terminal, or use scripts to make different apps work together.

For professional users such as developers and designers, these capabilities are important reasons for choosing Mac over iPad.

In fact, permission management on PCs is generally looser than on mobile phones.

On Windows, ordinary desktop software can usually directly read files that the current user has access to, without the need to apply for permissions one by one like mobile apps.

Even if there is a permission control mechanism such as UAC (User Account Control), its authorization method is still quite extensive——

As long as the user clicks "Yes", the software can obtain administrator rights and perform almost all operations including modifying system settings and installing other software.

Picture | Wikipedia

Although macOS imposes more restrictions on apps' data access through sandboxing and privacy permission mechanisms, it has always retained a way for some apps to obtain broader system permissions.

"Full Disk Access" is one of them.

Backup apps need to read the entire computer, security tools need to scan protected files, and some automation apps need to operate other applications.

For these scenarios, Apple allows users to proactively grant additional permissions.

Once full disk access is turned on, apps can bypass macOS restrictions on many files.

There are certainly risks to this approach, and the risks weren't there on day one. But in the era of traditional desktop apps, everyone can usually judge whether the authorization is reasonable based on the purpose of the application.

After all, it might seem okay that a junk-cleaning app needs to read your entire hard drive, but something is obviously wrong if even a calculator says it wants to access all of your files.

The problem is that in the AI ​​era, this method of judgment is no longer very useful.

Agent becomes more and more like a "person" who can act independently. It will "judge" what to do next, and even change its plan based on information acquired along the way.

You ask Codex to help organize a week's work data. It may need to read local files, find emails, or even operate other applications.

This is certainly a good thing. But the problem is that while this autonomy brings convenience, it also brings new security risks. It is difficult for you to know in advance which files it will open and what operations it will perform, and there is no guarantee that every step will comply with your intention.

What’s even more troublesome is that the Agent’s judgment may also be interfered with by external information, which is called “cue word injection”.

Attackers can insert malicious instructions into documents, emails, or web pages to induce the AI ​​to perform additional actions or even steal private data. The greater the system privileges obtained by the Agent, the greater the potential harm.

Even without deliberate inducement by external attackers, the Agent's own permission boundaries have begun to cause controversy. In September this year, Meta’s AI Agent application Muse was involved in a privacy controversy.

Technology columnist Jason Aten published an article alleging that Muse read his private iMessage conversations and pushed him writing suggestions without his expectation and without, he said, full disk access.

🔗 www.inc.com/jason-aten/metas-new-muse-ai-agent-read-my-private-messages-i-never-asked-it-to/91408202

Meta later denied this claim, saying that reading messages requires the user to actively enable full disk access and connect to Messages.

While both sides disagree on the authorization process, the dispute raises a question:

If the user allows the AI ​​to access the chat history, does that mean it can browse private conversations at will and even use the content for purposes other than the current task?

Vulnerabilities in the AI ​​App itself may also become a breakthrough for attackers to obtain higher privileges.

In September this year, macOS security researcher Patrick Wardle discovered a security vulnerability in the Mac client of ChatGPT.

An attacker only needs to run a malicious program without administrator rights on the computer to have the opportunity to impersonate a trusted component of ChatGPT, borrow its permissions to read private chat records, access browser sessions, and even access data originally protected by macOS privacy mechanisms.

In order to let AI work freely, manufacturers first put shackles on it

macOS 27.3 Beta 2’s new permission warning quickly sparked debate in the developer community.

Apple hopes that users will fully understand the risks before authorizing, but for some tools that need to access a large number of files, overly eye-catching warnings may also make users unnecessarily suspicious of the app.

After the new beta version was pushed, alin23, the developer of the Mac image optimization tool Clop, complained on Reddit.

He explained that Clop provides optional full disk access permissions to monitor folders and reduce interruptions from repeated authorizations; photo permissions are used to obtain full-size images. However, the warnings corresponding to these permissions will make users think that this picture tool wants to export messages, emails, and even view hidden photos.

In the same post, the developers of the file navigation tool Default Folder

He admitted that these permissions did bring the risks listed in the warning, but he was worried that the eye-catching red warning would scare away users, and was ready to add explanations to the app's boot process.

Users who support the new warnings argue that the risks should have been known to everyone.

Netizen staires said that at a time when Vibe Coding has enabled a large number of new applications to emerge, an App that looks beautiful and has full functions is no longer enough to reassure people; the so-called users will protect themselves by checking open source code is unrealistic for most people.

However, some users believe that Apple’s warning does not exaggerate the risks.

Even if Clop only wants to read pictures, full disk access may still allow it to access private data such as emails and chat history. Just because the developer has no intention of reading this information does not mean that the App does not have such capabilities.

On the other hand, the more pop-ups appear, the easier it is for users to become numb.

Anthropic once published a set of Claude Code's background data: In the early authorization mechanism, in order to prevent the AI ​​from accidentally deleting files or executing dangerous commands, a pop-up box was displayed for consent at each step of the operation. As a result, users approved approximately 93% of the requests.

Once entering a high-frequency confirmation state, people will develop "authorization fatigue" and often click permission without even looking at it.

Faced with this dilemma, the industry began to try another idea: to withdraw control to the bottom layer of the operating system. Instead of asking people every time a pop-up window appeared, it was better to draw a "cage" for AI in advance that it could not get out of.

On macOS, AI programming tools have already taken the lead in using the system’s native Seatbelt sandbox.

Last October, Claude Code introduced underlying isolation based on Seatbelt, which limits file system and network access at the same time: AI can freely read code, change files, and run tests in the designated project directory, but it cannot touch private files outside the directory, nor can it connect to external servers without permission.

Swipe left to view the original English image, source Anthropic

After adding this hard boundary, Claude Code's permission confirmation pop-ups have been reduced by 84% - because as long as the AI ​​does not cross the boundary, there is no need to disturb the user. OpenAI's Codex also adopts a similar idea on macOS, also using Seatbelt to restrict local files and network access.

Microsoft proposed Agent Workspace in Windows 11, which assigns independent accounts and restricted workspaces to Agents.

Picture | Microsoft

Whether Apple breaks down the switches at the system level, or AI tools build their own sandboxes at the bottom, all parties are trying to redefine the scope of AI activities.

But these increasingly tight boundaries are also quietly changing the experience of ordinary people using computers.

Apple says it’s “good for you,” but Macs are getting harder to use

On Mac, I often encounter errors when dragging files from WeChat to other apps.

After some attempts, I found that actually I just need to right-click, copy and paste, or save the file to the desktop first and then drag it in from Finder to upload normally.

Some users on GitHub also encountered the same problem: dragging files directly from WeChat to ChatGPT will fail to upload, but if they are copied and pasted, everything will work fine.

After checking it out, I realized that this was related to the sandbox isolation of macOS.

As early as the macOS 14 Sonoma period, Apple tightened the data isolation between different applications: the files of each App are locked in their own areas and cannot be read from each other casually.

Files received by WeChat are stored in its own directory by default. If the recipient does not obtain cross-application read permission according to the system's new rules, the system will block the file and only give the user a vague "Upload failed."

Picture | Apple WWDC23

As long as you copy or save the file to the desktop first, the file will leave the "small room" of WeChat and become an ordinary file on the desktop, and other apps can naturally read it smoothly.

The troublesome thing is that this restriction is calculated one by one according to the "target App".

You want to drag it into Chrome today, and you want to drag it into ChatGPT tomorrow. Every time you change a target App, you have to face its permissions or adaptation issues - it is impossible to turn on "Full Disk Access" for every App on your computer just to save trouble.

For ordinary people, the originally intuitive "drag and click" operation has become hindered everywhere due to security compromises.

Picture | 9to5Mac

Apple continues to raise walls, add pop-ups and prompts on macOS, and the starting point is indeed to protect user privacy. But in daily use, increasingly fine restrictions, frequent warnings, and inexplicable error reports ultimately leave users with inconvenience and confusion.

Picture | MacRumors

But for users, what they ultimately want is actually very simple: to allow AI to work on its own within a reasonable range, and for people to continue to use computers conveniently.

If you want to take care of these two things, Apple, and even the entire industry, obviously still have a long way to go in terms of the integration of security boundaries and daily experience.

Zheng Tingxu

Hardware editor

One more revision.

A total of 350 articles have been published

This article comes from the column

Next article2026-11-18 08:32

Morning Post|Huang Renxun expressed his willingness to return to NVIDIA/Li Feifei joined AMD and reported directly to Su Zifeng/Hongguo short drama to deal with portrait infringement

Previous article2026-11-17 16:51

All the unlimited usage of DeepSeek are here | with guide

Aifaner App

Love style, make the future within reach

Digital trendy brands focusing on tomorrow’s products

Advanced tools, prophets first, AIGC’s inspiration guide

A new era of car manufacturing, tomorrow’s travelers

Explore new lifestyles in the city and be your guide to tomorrow’s life

Copyright © Guangzhou Liwo Zhiyuan Investment Management Partnership (Limited Partnership) 2008 - 2026. Use of Aifaner website content for commercial purposes requires permission. Non-commercial use is licensed under CC BY-NC 4.0.

All content is made available under the CC BY-NC 4.0 for non-commercial use. Commercial use of this content is prohibited without explicit permission.