A single click can give an AI agent the keys to nearly everything on your Mac. Apple wants users to think twice before handing them over.
Starting with macOS 27.3, currently in beta 2, Apple is overhauling how users grant sensitive permissions to apps, especially the AI agents that have made themselves at home on Macs. The company is splitting, renaming and adding friction to the controls that govern whether an app can read every file on your machine or control other apps.
The changes arrive amid growing scrutiny of AI agents' access to sensitive data. A journalist claimed Meta's Muse app read private messages without consent, while Wired reported a security flaw involving ChatGPT's Mac app. Apple's approach is not to block these tools, but to make the risks of granting access harder to overlook.
"When you give an app Full Disk Access… it can bypass system protections and access all of the files on your Mac, including your messages, emails, photos and videos, data from other user accounts, and more," Apple warns in a new support document. "It can read, change, delete, and export this data without asking."
Here is what changes:
Permission area | What changes in macOS 27.3 |
|---|---|
| Permission settings | Device Control and Data Access is separated into System & App Protection Override and Accessibility |
| Permission requests | Users receive more explicit warnings about sensitive access |
| App automation | Users can specify which applications another app may control |
| Approval options | Supported app-control permissions offer Ask First or At Any Time |
| File access | Supported requests allow more granular folder selection |
| Activity visibility | A menu bar indicator shows when one app is controlling another |
| Existing permissions | Previously granted access remains in effect |
What it means for Mac users
The new flow is designed to give you pause. If an app requests full disk access, macOS now forces you to review exactly what you're enabling: "Read, Change, Delete, and Export All Your Data," "View Hidden and Recently Deleted Photos," and "Browse As You on Websites." Only after acknowledging those risks can you proceed.
For supported app-control permissions under System & App Protection Override, users can select which applications an automation tool may control and choose between "Ask First" and "At Any Time." A menu bar indicator provides additional visibility when one application controls another. These controls are distinct from Full Disk Access.
Apple is careful not to break existing workflows. Permissions already granted remain in effect, and known apps like backup tools and assistive technologies can request entitlements to preserve their previous experience. Developers, however, must update their onboarding and ensure their apps "gracefully handle having access to only a limited set of folders or apps."
Why AI agents make broad permissions riskier
This is not an anti-AI move. The changes are particularly relevant to users running local AI agents on Macs. The threat is more specific: agents that run on-device, take actions autonomously and inherit whatever permissions their parent app holds.
"Before granting an app such extraordinary access, users should understand that it puts all of the data on their Mac at risk," Apple wrote to developers.
Meta's David Singleton pushed back on the Muse incident, saying the app can read message content "only if macOS system-level 'Full Disk Access' has been granted and the Messages connector is enabled." The distinction highlights a broader security concern: an AI agent may be able to access sensitive information through permissions users previously approved, even when they did not anticipate how extensively those permissions could be used.
What security teams should check before macOS 27.3
Apple's changes give users more information before approving sensitive permissions, but they do not eliminate the risks associated with applications that already have broad access.
For enterprise security teams, the bigger concern is whether AI agents and automation tools have more privileges than their tasks require. An agent with extensive permissions could potentially access confidential documents, interact with business applications, or expose sensitive information if compromised.
Security administrators should review several areas before deploying macOS 27.3:
- Audit Full Disk Access: Identify applications with existing permissions and determine whether they still need unrestricted access to files.
- Review AI agents and automation tools: Check which applications can control other apps, interact with sensitive data, or perform actions without repeated approval.
- Apply least-privilege principles: Where supported, limit access to specific folders or applications instead of granting broader permissions.
- Check existing approvals: Previously granted permissions remain active, so administrators should not assume the update automatically removes excessive access.
- Test business-critical applications: Confirm that backup software, accessibility tools, and enterprise automation workflows continue functioning under the revised controls.
- Review monitoring capabilities: Determine whether existing endpoint security tools can detect suspicious agent activity after permissions have been granted.
The changes also expose a limitation of permission-based security: consent does not guarantee safe behavior. A legitimate application can still misuse its access, whether through a compromised component, malicious instructions, or unintended automation.
For organizations adopting autonomous AI tools, Apple's new controls should be treated as another layer of defense rather than a substitute for endpoint monitoring, application governance, and regular permission reviews.
Read more: Stronger Mac permissions are only one part of securing autonomous AI. Learn why AI agent security requires more than instructions and how organizations can enforce access boundaries, limit excessive privileges, and test containment before deployment.





