Apple Moves to Lock Down Full Disk Access, Citing AI Agent Risks
Apple announced on October 2 that Full Disk Access will require much more explicit user consent going forward, citing AI agents specifically: 'as AI agents become increasingly capable and autonomous, the risks associated with this level of access will grow substantially.'
Apple is tightening Full Disk Access. In a developer announcement on October 2, the company said it will introduce additional controls so that users can grant “this extraordinary level of access” only through very explicit user action going forward, and the reasoning names the pressure point directly: “As AI agents become increasingly capable and autonomous, the risks associated with this level of access will grow substantially.” The announcement notes that Full Disk Access “largely sidesteps” privacy controls, was originally designed for backup software, and is now being used in ways that expose “files, mail, messages, and even browsing history” without users’ full understanding, adding that for communication apps the privacy of the people users talk to is compromised too.
This Is the Muse Aftermath, Written by the Vendor
Read the announcement against the past two weeks and the target is unmistakable. A professor’s controlled test found Meta’s Muse synced roughly 187,000 lines of his Messages database with Full Disk Access off and Messages permission never granted, and the researcher who disclosed Muse’s auth-token zero-day pointed at the same design flaw: agents treating elevated access as ambient rather than granted. Apple’s language, that some developers use Full Disk Access in ways that “could put users at risk,” is the platform vendor’s response to exactly that behavior pattern. Platform privacy promises were never designed for software whose value proposition is ingesting everything it can read, and Apple is now re-fencing the boundary in public.
The Consent Problem Is Harder Than a Dialog
The technical challenge Apple has set itself is that Full Disk Access consent was always a weak contract. Users grant it once, in System Settings, to a backup app, and the grant then covers every future behavior of every process the app spawns, including agents that didn’t exist when the checkbox was ticked. “Very explicit user action” implies context-bound consent: per-operation prompts, time-limited grants, or scoped access that names what will be read (Messages, not the whole disk). None of the concrete mechanics are announced yet, which is the gap between this document and a shipped feature. The announcement also hints at the multi-party problem this blog flagged in the Muse case: communication apps expose the people you talk to, so consent from the user installing the agent cannot cover the message content of everyone they correspond with.
What Developers and Agent Builders Should Do Now
The practical guidance for anyone shipping agent software on macOS: stop treating Full Disk Access as an acceptable integration path for reading user content, and migrate to scoped APIs (Apple’s suggestion in the announcement) before the new consent flows make the broad grant both harder to get and a trust liability to display. For agent products specifically, this is a design deadline, not a paperwork deadline. The platforms are converging on the same lesson from different directions: Nvidia moved agent containment into network silicon, Google ships Gemini 4 Argon to vetted defenders before the public, and now Apple is making elevated disk access an explicit, informed, probably per-use decision. Ambient access is being designed out of the stack, layer by layer.
What to Watch
Three things. First, the concrete mechanism when Apple ships it: per-file scopes, per-agent prompts, or a new entitlement category will determine how much existing agent software breaks. Second, whether Meta responds by moving Muse onto scoped APIs or fights for its current access, which will reveal whether the ingest-everything design can survive platform enforcement. Third, whether other OS vendors follow: the Muse incidents proved a consumer agent will read what it can, and every platform with a permissions model is now deciding whether its own version of Full Disk Access is an AI-era liability. Apple moved first; the answer to whether this was overcorrection arrives with the first agent that finds its data genuinely walled off.
Get Insanely Good at AI
The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.
Keep Reading
How to Build Advanced AI Agents with OpenClaw v2026
Learn to master OpenClaw v2026.3.22 by configuring reasoning files, integrating ClawHub skills, and deploying secure agent sandboxes.
Meta's Muse Synced 187,000 Lines of Messages With Permissions Off
A professor's test of Meta's Muse found it pulled roughly 187,000 lines from his iPhone Messages database within a day, with Full Disk Access off and Messages permission never granted, per AppleInsider's report on his account.
ChatGPT Wallet lets OpenAI's dots pay with network tokens, no OpenAI cut at launch
OpenAI launched ChatGPT Wallet on November 13, 2026, letting its always-on dots agents buy things with scoped Visa, Mastercard and Amex tokens. Here is how the payment path works, what merchants must change, and why it starts so small.
Claude Sheets, Whiteboards and Meetings: Anthropic builds its own office suite
On November 10 Anthropic launched Claude Sheets and Whiteboards in beta, made Claude Data generally available, added interactive connectors from Adobe and Atlassian, and began a Claude Meetings beta that puts Claude into Google Meet calls.
DeepMind's AlphaProof Code proved Signal's security on the compiled binary: 6.5M lines of Lean, no human-written proofs
Google DeepMind's AlphaProof Code used AI agents to prove forward secrecy, post-compromise security and authentication for libsignal's AArch64 machine code in four weeks. Here is what was proved, what was assumed, and what the agents got wrong along the way.