deniz.in

Markets

Weather

Loading weather

· via TechCrunch

Apple overhauls Mac permission controls in macOS 27.3 to rein in AI agent access

Apple is splitting device control and data access permissions, adding per-request approval prompts and a transparency indicator in macOS 27.3, responding to AI agents that demand full disk access.

Apple overhauls Mac permission controls in macOS 27.3 to rein in AI agent access

New permission model arrives with macOS 27.3

Apple has detailed a rework of the Mac's security permission system, shipping with macOS 27.3, that responds to a wave of AI agents asking users for sweeping access to their machines. As TechCrunch reports, the company committed to revising these settings in October and has now published a developer-facing blog post explaining the changes, warning that before an app is granted such broad access, users should understand it places all the data on their Mac at risk.

The most structural change is a split. What was previously the combined Device Control and Data Access setting becomes two separate options: System & App Protection Override and Accessibility. Apple's intent is to make it unmistakable that enabling the first option means overriding the operating system's built-in protections rather than tweaking a convenience setting.

Granting permissions is also being consolidated into a single flow in which the user selects the requesting app from a list. Guidance will appear at the moment of enablement, encouraging users to scope System & App Protection Override to only the apps that genuinely need it, with similar prompts shown when Full Disk Access is turned on for folder access. Apple also promises greater visibility into the range of data an app would see or control if full permissions were granted.

The backdrop: agents with broad reach

Unlike chatbots that respond to prompts, agents run locally and act on the user's behalf across the operating system, files and other applications, which is precisely what makes their permission requests so consequential. According to TechCrunch, two incidents shaped Apple's thinking: a journalist's claim that Meta's Muse app read private messages without consent, and a Wired report, cited by TechCrunch, on a flaw in the Mac version of ChatGPT that could have let attackers reach sensitive data.

There is also a commercial tension at play. As TechCrunch notes, Apple has been selling Mac minis and Mac Studio machines to early adopters who buy them specifically to run local AI models, so it has little interest in blocking agents outright. The new controls are calibrated accordingly: they do not stop users from installing agents or any other app that requests elevated access.

Finer-grained consent and visibility

Two additions target the moment-to-moment behavior of agents. Users will be able to configure an app to ask for permission every single time it needs access, rather than working from a standing approval. TechCrunch observes this may prove tedious and could stand in the way of true automation, but it gives users a way to experiment with local AI without handing over blanket access to their data.

A new transparency indicator will also appear whenever an app is controlling another app through the System & App Protection Override option, so users can see when an agent is actively driving other software on their machine.

Developer guidance and exceptions

Apple is advising developers to rework their onboarding flows around the new controls, and to design their apps so they keep working even when their reach is restricted to only some folders or applications. There is an escape hatch as well: developers whose apps rely on Full Disk Access for disk backup or to deliver assistive technologies can request what Apple calls entitlements, preserving the behavior their apps had before the update.

Why it matters

Agentic AI inverts the assumption behind traditional app permissions: an agent's whole purpose is to act across your system, so it requests exactly the kind of blanket access that existing consent dialogs were never designed to police. Apple's response is notable because it does not pick a fight with the category. Instead of banning agents or capping their capabilities, macOS 27.3 tries to make the trade-offs legible, separating protection overrides from accessibility, exposing what a grant actually covers, and letting users demand a fresh confirmation for each action. The approach will create friction, and builders of automation tools will feel it most, but it sets a reference point for how platform owners can accommodate agentic software without asking users to sign away their machines wholesale.

  • #apple
  • #macos
  • #security
  • #privacy
  • #ai-agents

Related posts