Who Owns the Permission on a Managed Endpoint

On October 2, Apple said it is tightening macOS Full Disk Access controls and gave a reason that no earlier macOS privacy change had cited: new risks from AI agents. Full Disk Access is the permission that lets an application read files across the whole system, including mail, messages, browser data, and other applications' stores. A desktop agent that can read everything on a laptop is useful for exactly the reason it is dangerous. Apple's decision is a platform vendor acknowledging that the permission model built for human-driven apps needs revisiting for software that acts on its own.

The same week, OpenAI's Dots, the persistent agent product launched September 30 with dedicated cloud VMs and browser access, became available for enterprise deployments managed through Microsoft's Agent 365 security controls. Microsoft describes Agent 365 as a control plane for observing, governing, and securing agents, with Copilot Studio agent support planned for October, a date I could not confirm against a primary source. Two vendors, in one week, announced controls over what agents may do, from two different layers of the stack.

For an organization managing a fleet with Intune and Automox, that raises an architecture question. When an agent asks to read a protected folder on a managed Mac, which layer decides?

There are at least four candidates. The operating system holds the first decision, through its own privacy prompts and the permission database behind them. The device management layer holds the second: an MDM-delivered privacy preferences profile can pre-approve certain permission categories for specific applications and can pre-deny others, which means an administrator has already answered some of the questions before a user ever sees a prompt. The agent platform's own gating holds the third, such as Dots' routing of account-affecting actions through auto-review and user-defined rules. The identity layer holds the fourth: Entra conditional access, workload identity, and the scopes of the tokens the agent carries decide what it can reach once it leaves the device.

Each layer is reasonable alone, and the interesting failures live where they disagree. An MDM profile that grants Full Disk Access to an agent application, for instance, makes the operating system layer silent. The agent then holds whatever the agent platform's own gate allows, and the identity layer restricts only what happens off the device. A permission that looks tightly controlled at three layers can be wide open at the fourth, and nobody notices because each team's console shows green.

Two failure shapes seem most likely. In the first, a layer is configured permissively because another layer is assumed to be restrictive, and the assumption was never written down. The endpoint team approves broad file access because the agent platform has a review gate. The agent platform team relaxes the gate for read-only actions because the endpoint team controls file access. Each side reasonably believes the other holds the line. In the second, two layers hold contradictory policies, and the effective permission becomes whichever layer evaluates first, a behavior that can change with an operating system update.

What would help is an explicit statement of which layer is authoritative for which question, written before the first pilot. My own inclination is to assign each class of decision to the layer that has the best information. File and device access decisions belong to the endpoint layers, because only they see what is on the machine. Decisions about what the agent may do inside business applications belong to the identity layer, because that is where scopes and entitlements already live. Decisions about which individual actions need a human in the loop belong to the agent platform, because only it sees the action in context. In that arrangement the other layers can only restrict further.

The part I find most underappreciated is auditability. A permission decision spread across four consoles produces four partial logs, and answering the question of what the agent could have accessed on a given day requires joining all of them. A subpoena or an incident review asks that question in exactly that form. An organization that can answer it from one place has done a layer of design work that most deployments have not yet attempted.

Apple's announcement matters beyond Macs because of what it signals about timing. The platform vendors have begun to tighten defaults before agent deployments reach scale. The timing leaves a short window in which ownership is still cheap to settle, because permissions granted by default and by habit are expensive to unwind.


Sources: TechCrunch, "Apple says it's tightening macOS Full Disk Access controls due to new risks from AI agents" (Oct 2, 2026); Microsoft 365 roadmap items on Agent 365 and Copilot Studio as summarized in AI Industry Digests (Oct 2-12, 2026); OpenAI Dots launch coverage (Sept 30, 2026). The Copilot Studio support date was not confirmed against a primary source.