Apple moves to curb Full Disk Access as AI agents ask for the whole Mac
Apple said on 2 October that it will add new controls to macOS Full Disk Access, the permission that lets an app read files, mail, messages and browsing history. Developers began using it in ways the company says could put users at risk.

The announcement came in a post on Apple's developer website. The company attached no timeline, no technical detail and no list of which apps it has in mind. 9to5Mac reported the post on 2 October.
For Mac users, granting an app blanket access to the disk should soon require a deliberate act. Today that takes a click buried in a settings pane. What is still unknown is what that act looks like. Apple has not said whether existing permissions get revoked on update, or which apps lose the ability to work as they do now. The gap between a promise of controls and a shipped control is where most of this story still sits. Nobody outside Apple can yet say how wide it will be.
Full Disk Access was built for backup software. Apple's post says the permission "largely sidesteps" the controls it uses to protect private data. It adds that "some developers are using Full Disk Access in ways that could put users at risk, exposing everything on their systems, including files, mail, messages, and even browsing history, without users' full knowledge and understanding." The same post warns that for messaging apps the exposure reaches the people a user talks to.
Why the permission became a problem now
The timing is not subtle. Always-on assistants such as Meta Muse and OpenAI's Dots ask for broad access to personal data. 9to5Mac notes that Muse has already produced what it calls horror stories. An agent that can read every file on a laptop is more useful and far harder to bound.
Apple framed the change around that trend. "As AI agents become increasingly capable and autonomous, the risks associated with this level of access will grow substantially," the company wrote. It added that it wants users to understand the risks before granting access.
Agents are also arriving at the browser from the other direction. A developer post by Marko Denic on 3 October describes WebMCP, a Chrome experiment. It lets a page expose its forms and functions to an agent as named tools with typed inputs, so the agent calls a function instead of guessing at pixels. Three HTML attributes on an existing form are enough. Browsers that do not support WebMCP ignore them.
Denic's example is a newsletter search form, tagged with a tool name and a description. He recommends leaving the auto-submit attribute off anything that sends, buys or deletes, so a person still confirms. The trial currently runs on Chrome 149 and needs an origin trial token. Most visitors will not see it yet.
The measurement problem behind the controls
Permission prompts assume someone is watching. A paper by Emre Turan, posted to arXiv and dated 8 June, argues that the human approval gate is the easy part. The judgement behind it is not. On 125 hand-labelled agent actions, reviewers agreed only moderately on what counted as risky, with a Fleiss' kappa of 0.52.
Turan models the reviewer as tiring as escalations pile up. He finds that safety forms an inverted U as the escalation rate rises. Past a point, more human oversight makes a system less safe, and the best setting escalates below full escalation. He is explicit that the mechanisms are prior art, citing fatigue-aware learning-to-defer and cost-sensitive deferral work. He also says the flooding attack and the inverted-U are modelling results meant to motivate a human study.
"The gate is the easy part; the hard part is the judgment," the paper's abstract states.
Microsoft and Hugging Face published a benchmark on 3 October that measures something adjacent: what an agent leaves behind. ThinkingBox runs agents against isolated tool sessions and grades the final state of the backend rather than the reply. Across 507 stateful business workflows, each run 20 times, the common-set ablation covered 121,680 valid trials across 12 models. 79,853 attempts failed the executable checks.
Of those failures, 67.24% ended cleanly, called a state-changing tool and reported no error. The checks still found wrong field values in 77.61% of them, extra side effects in 43.30% and missing required effects in 25.36%. The authors sum it up with a line worth keeping: a trajectory is a claim, database state is the evidence.
What builders are doing instead of asking
The tooling around agents is shifting toward giving them their own credentials and their own surfaces rather than a user's. A post from AgentID on 4 October lays out the case for issuing an agent its own inbox and sign-in key. Revoking it then means deleting that key rather than changing your own password.
On the desktop, more-space is a small app that puts a terminal on an endless chalkboard. It lets the agent draw its plan, its progress and its questions there, with an avatar for company. It installs nothing into the project.
easl, described by Tim Waldin on 4 October, takes a similar idea further. It gives one infinite board per repository, with agents in real terminals beside the diagrams and pages they produce. A keyboard shortcut adds whatever you point at to the next prompt. Waldin builds the argument on a post by Andrej Karpathy, who said we will spend a lot more time trying to understand what language models output.
Luke Wroblewski makes the design case on the same day. If agents write most of the front-end, a design system becomes a steering layer. It brings tokens and an AGENTS.md file that every agent reads before it adds a button. On his projects the code, not Figma, is the source of truth.
None of this fixes the permission Apple is worried about. A design system keeps an agent consistent. It does not decide whether that agent should be able to read your mail.
The oversight gap Apple's controls will not close
Vendors are selling governance as the answer. Kimchi, launched on 4 October, routes every agent request through one enforcement layer. It sets hard spending caps per user, team, API key or model, and offers to run inside a customer's own cloud. Spawned, also out on 4 October, provisions infrastructure into a customer's AWS, GCP or Azure account. It pitches itself as the way to see what agents built.
Those products address cost and configuration. They do not address judgement, which is where the research keeps landing. Sandeep Archarya's post on agent drift, published 3 October, asks the question that catches teams out: if nobody changed the agent, why did its behaviour change? The answer is that models, tools, data and other agents all move underneath it. Monitoring the model alone misses the shift.
Apple's answer is narrower and more concrete. It will make the permission harder to give. Whether that protects anyone depends on details the company has not published. It also depends on whether the agents asking for the disk can do their work with less.
Sources
11- 01Apple says it's tightening macOS privacy controls amid the rise of AI agentsEN
- 02Let AI agents use your site with 3 HTML attributesEN
- 03Oversight Has Capacity: Calibrating Agent Guards to a Subjective Fatiguing HumanEN
- 04The Agent Said It Was Done. The Database Disagreed.EN
- 05Give Your AI Agent Its Own IdentityEN
- 06Chalkboard and Avatar for AI AgentsEN
- 07Easl.sh – your agents diagrams need somewhere to liveEN
- 08Design Systems for AI AgentsEN
- 09Pi-based coding agent with hard budget capsEN
- 10Spawned: AIO cloud platform for humans and agentsEN
- 11Understanding Agent DriftEN
All figures and quotations in this text come from the sources listed below.
Content prepared by the editorial team with AI assistance.
Comments
0- No comments yet — be the first.