The Permission Is the Policy

· 5 min read agentspermissionsenterprise-security

Navy and amber graphic with the headline The Permission Is the Policy

On October 2, Apple announced it is tightening macOS Full Disk Access. The reason it gave is worth reading twice: AI agents have increased “the risks associated with this level of access,” and “as AI agents become increasingly capable and autonomous, the risks associated with this level of access will grow substantially.” (TechCrunch)

Full Disk Access is one of the oldest privacy permissions on the Mac. It lets an app read files, mail, messages, and browsing history. Apple built it so backup software could work. That permission was fine for a backup tool, because a backup tool reads quietly in the background and never thinks about what it read.

The trigger for Apple’s announcement was a report that shows exactly what changed. Inc. columnist Jason Aten said Meta’s Muse agent on his Mac synced more than 187,000 rows of his private iMessages after he says he declined to grant it message access. Meta disputed the claim. A separate Wired report documented a flaw in the ChatGPT Mac app that could have exposed sensitive data. I am not adjudicating either incident. What matters is Apple’s judgment: the same permission means something different when the holder is an agent.

The grant changed. The checkbox did not.

A backup tool with Full Disk Access can see everything on your disk. An agent with Full Disk Access can see everything on your disk, interpret what it finds, connect it to another system, and take the next step. The permission did not get broader. The holder got more capable. That is the entire story, and it is why Apple is moving to require what it calls “very explicit user action” before this level of access can be granted going forward.

This is the same problem I have been circling all month, arriving through the platform’s own front door. In We Need SOX for AI I argued for an inventory of every agent in production, with an owner and a control set. In Always-On Agents Need Declared Boundaries I proposed six declared terms for every agent: the surfaces it may touch, the actions it may take, the spend it may incur, the data it may reach, the principals it answers to, and how it gets revoked. The permission grant is where all six of those terms collapse into a single click. A user grants it once, to do one job, and it becomes the standing authority for every job the agent takes on afterward.

The enterprise version of this is not a Mac permission. It is the service account the agent runs under, the OAuth scope it was granted at setup, the shared mailbox it can read, the admin API key pasted into a config file during a pilot. Every one of those is a Full Disk Access of its own: a broad grant made once, for a narrow purpose, that the agent then carries everywhere.

Ask four questions before any desktop agent ships

I wrote this month about an Amazon AI bot that took down its own cloud, and the postmortem called it “user error, specifically misconfigured access controls,” because the engineer’s permissions were broader than expected. The bot was incidental. The permission was the blast radius. The rule has not changed since: restrict what the agent can touch, limit the scopes it holds, require human approval where the stakes are real.

Before a desktop agent or an internal agent with system access ships in your org, make the security review answer four questions, in writing, and file the answers where the agent’s other controls live:

  1. What can it see? Name the data sources, not “the filesystem” or “email.” Files in which directories, which mailboxes, which message stores. If you cannot name them, you do not know.
  2. What can it send? Read access is half the story. An agent that can read the contract and also send email can do more damage than one that can only read. Separate the read scope from the write scope, always.
  3. What can it change? The destructive surface: file writes, deletions, API calls that mutate state, money movement, messages sent on someone’s behalf. Each one needs its own justification.
  4. Who can reconstruct what it did? The action log. Every agent decision and tool call recorded, attributable to the agent and its task, reviewable after the fact. If the agent acts and nobody can reconstruct the actions, you have an operator with no audit trail.

A checkbox and a privacy policy will not answer these. Neither will the vendor’s marketing page. The answers are your policy, and the permission you grant is how the policy is enforced at runtime.

The grant is the policy document

I have a control-grammar concept in my own work where a higher authority defines the envelope within which a lower one operates, and it may delegate any portion of that envelope downward but never beyond it. The principle is old and unglamorous: least privilege, written down. What Apple is doing is the platform version of the same discipline. It is not banning the permission. It is forcing the grant to be an explicit, informed decision, because the grant is where authority is actually defined.

Apply that to every agent you deploy. The permission is not a prerequisite you click through to get to the demo. It is the policy document that governs everything the agent will do after. Write it like one.