Apple separates disk and automation permissions in macOS to contain AI agent risks
Apple detailed on Monday the security changes promised in October for macOS, splitting full disk access into two permissions and requiring more transparency from apps that run AI agents on the Mac.

- Reading
- 5 min
Apple detailed on Monday the security changes promised in October for macOS, splitting full disk access into two permissions and requiring more transparency from apps that run AI agents on the Mac.
What Apple changed
Apple announced on Monday (11/16) the macOS security changes promised in October 2026, aimed at containing the risks AI agents pose to user data. The changes arrive with macOS 27.3 and split into two the permission that today is granted all at once: what Apple now calls System & App Protection Override and Accessibility (source: TechCrunch).
Today, granting "full disk access" to an app on the Mac is essentially all-or-nothing: the app gains unrestricted access to files and folders, as well as the ability to control other apps via automation. With the split, the user chooses whether to grant only data access (Accessibility) or also grant control over other apps via automation (System & App Protection Override), an option Apple makes clear overrides native system protections.
Before granting an app such extraordinary access, users should understand that it puts all of the data on their Mac at risk.
Apple, in a statement published on the developer blog
Why now: two incidents that pressured Apple
The review didn't happen in a vacuum. Apple had been under pressure following two cases cited in TechCrunch's report: a journalist reported that Meta's Muse app had allegedly read private messages without consent; and Wired published a flaw in the ChatGPT app for Mac that could have allowed hackers to access sensitive user data.
At the same time, Apple has been profiting from selling Mac mini and Mac Studio units to early adopters who buy the machines specifically to run AI models locally. The company needs to balance two opposing interests: encouraging this local AI market on its own hardware while, at the same time, not letting the operating system become an open door for agents with unrestricted access.
Unlike a chatbot that only answers questions, an agent runs on the user's device and can act on its own on the operating system, files, and other apps on behalf of whoever installed it. It's this autonomous behavior, not AI itself, that prompted the permissions review.
What changes in practice
In addition to separating the two permissions, Apple packaged five other changes into the same security bundle:
- A single grant flow: the user now selects the app from a list, instead of dealing with dialog boxes scattered across different preference screens;
- Active guidance for the user: when enabling System & App Protection Override or Full Disk Access, macOS now explicitly recommends granting access only to apps that truly need it;
- More visibility into scope: Apple promises to more clearly show the "wide range of data" an app gains control over when it receives full permissions;
- Per-event permission: the user can configure an app to request authorization every time it needs access, instead of a permanent grant, a setting that, in TechCrunch's assessment, can be tedious and represent an obstacle for end-to-end automations, but gives fine-grained control to those who want to experiment with local AI without giving it carte blanche;
- Transparency indicator: a visual signal now appears when an app is controlling another via System & App Protection Override.
None of these changes prevent installing or running AI agents and other apps that request high levels of access. What changes is the amount of information and granular control the user has over the decision: the model remains opt-in, it just becomes more explicit.
What this means for developers in the Apple ecosystem
For those building apps or agents that depend on disk access, the Accessibility API, or cross-app automation on the Mac, Apple's direct recommendation is to review onboarding. The company asks that apps learn to gracefully handle limited access to a restricted set of folders or apps, instead of assuming they'll receive full disk access right away.
In practice, this pushes product teams to rethink flows such as:
- Detecting and communicating when the app only has partial access, without breaking the experience;
- Reducing the permission scope requested on first contact, leaving additional requests for when a specific feature is actually used;
- Testing the app under per-event permission mode, since some users will likely opt for this more restrictive setting instead of a permanent grant.
Apps that depend on Full Disk Access for disk backup or to provide assistive technology have an option: they can request exceptions (what Apple calls "entitlements") to keep the same functionality they had before macOS 27.3. This applies, for example, to automation tools that historically relied on unrestricted access via AppleScript or the Accessibility API, a common pattern among AI agents that control other apps on the desktop.
For those building local agents, whether integrating models via Ollama or LM Studio, or packaging an AI wrapper that automates tasks in Finder, Mail, or third-party apps, the practical lesson is: permission UX has stopped being a configuration detail and has become part of the product. An agent that today requests full disk access generically, without explaining why it needs each folder, risks being rejected by more cautious users as soon as the new interface makes the request more visible.
What remains open
TechCrunch doesn't cite an exact release date for macOS 27.3, nor a beta timeline for developers to test the new permission screens before the general public. There's also no detail on how the entitlement request process will work in practice (who evaluates it, how long it takes, what criteria apply), nor whether apps already published on the Mac App Store will need recertification once the update arrives.
What's clear, however, is the signal Apple is sending: the platform will continue to allow AI agents with deep system access, but the price for that is more explicit information for users and more design responsibility for those who build these agents.
Translated from the Brazilian Portuguese original · Read the original
Architecture separates reasoning and execution for AI agents in critical operations
An architecture pattern described by InfoQ shows how operations platforms in regulated industries let the model propose actions, but never execute them without passing through a policy gateway outside the AI's control.






