Apple is tightening Mac security controls to manage the data exposure risks posed by local AI agents, introducing granular permission settings in macOS 27.3.

Unlike AI chatbots that answer questions, agents execute tasks directly on a user's device—accessing files, controlling applications, and operating across the operating system. That scope creates a new threat model: an agent with full disk access can read every private message, financial record, and document on the machine.

The company warned developers in a blog post that granting applications "extraordinary access" puts "all of the data on their Mac at risk." The warning follows two high-profile incidents: a journalist's report that Meta's Muse app accessed private messages without consent, and a flaw in ChatGPT's Mac app that could have exposed sensitive data.

Apple's approach splits permissions into two distinct categories. The first—System App Protection Override—now requires explicit opt-in and no longer bundles full disk access. The second, Accessibility, remains separate. The goal is to clarify that enabling the first option bypasses built-in security protections, not just enables automation.

When users grant System App Protection Override or Full Disk Access, they will select the specific application from a list rather than confirm a broad permission prompt. Apple will also provide guidance encouraging users to grant access only when necessary.

The company is adding a transparency indicator that alerts users in real time when an application controls another application through the System App Protection Override setting.

Crucially, Apple is introducing per-action permission requests for AI agents. Rather than grant unrestricted access once, users can configure applications to request permission each time access is needed. This friction slows automation but lets users experiment with local AI without opening their entire system.

These changes don't prevent users from running AI agents or other high-access applications. They simply force more deliberate choices about which applications get what data.

The move addresses a business tension for Apple. The company sells Mac mini and Mac Studio machines specifically to AI enthusiasts running local models—a growing revenue stream. But unrestricted agent access threatens user data and Apple's privacy-first brand positioning. The permissions redesign threads that needle: keeping the hardware attractive to developers while making the user experience less risky for ordinary customers.