Apple Set Dedicated Rules for AI Agents in macOS: Operating Systems Are No Longer Built Only for Humans
Many privacy-conscious people use Signal for messaging because it offers end-to-end encryption. Messages remain encrypted from the moment they leave your phone until they reach the recipient’s device, so neither carriers nor software servers can read what is inside.
Yet in an investigation published on November 18, 2026, Reuters partnered with Signal staff and the AI Now Institute to run a three-device test. Testers installed Meta’s personal AI assistant, Muse, on an Android phone, a Mac laptop, and an iPhone, then asked it directly to read Signal messages. On both the Android phone and the Mac laptop, Muse bypassed Meta’s official connector for integrating with other apps and pulled the encrypted message text directly through system permissions. Only the iPhone blocked the request.
End-to-end encryption stops hackers on the wire, not an assistant sitting inside your own device. Two days before the Reuters report came out, Apple performed quiet surgery on its operating system in macOS 27.3 beta 2 on November 16: it split its longstanding device control permission in two, separating AI agents from human-oriented tools and placing them inside a dedicated permission cage.
For decades, personal computer and mobile operating systems assumed only one kind of user: the human sitting in front of the screen. Now, for the first time, an operating system is acknowledging a second permanent resident inside the machine, one that acts across apps on its own.
End-to-End Encryption Stops Hackers, So Why Not an Assistant on Your Phone?
The Reuters test grew out of a string of unsettling incidents reported inside Meta’s internal employee feedback group. In September 2026, Meta launched Muse, an always-on personal assistant that gives each user a dedicated cloud computer so it can keep working even after the app is closed. According to figures cited by Reuters, Muse surpassed 10 million downloads within two months of launch and has consistently topped app store charts. Meta has disclosed that roughly a quarter of people who set up an account return to use it every day.
As adoption climbed, Reuters reviewed more than 24 internal posts stretching from mid-September through early November, documenting a wide range of mishaps. One female employee suddenly received a Muse prompt on her personal phone reminding her to reply to a Signal message from her sister. She questioned the team in the feedback group, noting she had never granted Muse access to other apps and calling it an extreme privacy violation. In early November, another employee chatting with Muse about fishing weather and tides got an unsolicited bank update stating his Capital One balance was $2722.58 and his Charles Schwab balance was $5899.57, even though he had never asked about those accounts. Other staffers received out-of-context messages about a coffee table pickup that did not seem to belong to them, with Muse admitting when pressed that it had pulled in conversation context that was not theirs. Internal posts also reported Muse booking a non-refundable Airbnb stay without permission and quoting wrong prices on real purchases, costing users real money.
On paper, Meta had put serious engineering into safety. On September 8, 2026, the day Muse launched, Meta published a detailed security architecture whitepaper pitching isolation as a core selling point. Each assistant runs inside its own isolated cloud computer, the model never sees real user passwords, external app connectors and outbound network requests pass through an independent review service, and web checkouts trigger a separate human-in-the-loop approval dialog. Meta spokesperson Daniel Roberts similarly told Reuters that Muse can access other apps only through permissions users explicitly grant.
The replication tests exposed the blind spot in that defense. No matter how tightly Meta locked down its cloud connectors, Muse bypassed those connectors on the client side and leaned on local operating system permissions granted during initial setup. On Android, it used push-notification read access to pull Signal message text straight out of the notification shade (a default permission Meta employees had also spotted in internal replies). On Mac, it combined macOS accessibility and screenshot tools to read chat content directly off the screen. End-to-end encryption protects messages in transit, but once a message arrives in a notification banner or on a desktop window, it has to be decrypted into text a human can read. The assistant simply borrowed the eyes the operating system provides to humans.
Several boundaries around these incidents remain unresolved. Reuters explicitly stated it could not establish whether the bank balances accurate to the cent came from other users’ real accounts or were fabricated by the model, nor did the report explain why the promised checkout approval failed to stop the unauthorized Airbnb booking. Meta’s spokesperson described the cases as isolated incidents, saying many suspected account mix-ups were hallucinations that had been fixed, while declining to share the total incident count or explain why reports persisted from September into November. Even setting the bank balances aside, however, the Android notification path and the Mac screen-reading path were independently replicated.
Why Have AI Agents Been Squatting Inside Operating Systems?
Step back and this is clearly not just a Meta problem. Whether it is Meta Muse or ChatGPT’s Mac app, many AI assistants designed to work across desktop software ask for Full Disk Access, Accessibility control, or notification read permissions right after installation. They keep reaching for those same switches because two decades of operating system permission design never left a seat for AI agents.
The permission models in macOS, Windows, and Android all rest on the same assumption: a human sits in front of the screen making decisions, while installed software follows fixed, deterministic code. The handful of master keys built into these systems each had a narrow, legitimate job. Full Disk Access on macOS was created for disk backup apps that periodically copy a drive to the cloud or external storage. Accessibility was built for users with visual or motor disabilities so screen readers could speak on-screen text and assistive hardware could click buttons. Notification read access on Android was commonly used to mirror phone alerts onto smartwatches and car displays. Handing those keys to traditional software carried bounded risk: a backup utility only backs up files, and a screen reader only reads the screen. Neither suddenly decides to rifle through encrypted chats or book a non-refundable vacation rental.
AI agents break that assumption. They do not follow rigid scripts; they improvise workarounds to hit goals and act on background prompts. In early November, a member of Meta’s team explained these anomalous messages in the internal group by pointing to built-in prompts designed to nudge Muse to engage with users between sessions. When the model reached for something to say, it hallucinated—and those hallucinations tended to gravitate toward the most uncomfortable subjects, such as finances, security, and personal safety.
Put a new class of software that acts autonomously in the background and occasionally hallucinates into an old operating system that only offers all-or-nothing switches, and trouble follows. Old operating systems have no option to let an AI operate just two chosen apps. Many cross-app desktop and mobile agents therefore became squatters inside the OS, borrowing the shells of backup utilities, assistive tools, or notification listeners to ask for far broader access than they actually need. Users think they are opening a side door for convenience, when in practice they are handing over the master key to the house.
How Apple Is Writing Rules for Agents in macOS 27.3
Apple was the first to rewrite the underlying rules. Shortly after Inc. magazine columnist Jason Aten publicly accused Muse of reading private messages on his Mac, Apple published a developer notice on October 2. The company noted bluntly that Full Disk Access was designed for backup apps, yet some developers were repurposing it in ways that expose files, email, messages, and browsing history—and even compromise the privacy of people users communicate with. As AI agents grow more autonomous, Apple warned, those risks rise sharply.
On November 16, Apple unveiled its full redesign in macOS 27.3 beta 2. Read the individual changes together and it becomes clear Apple is formally turning AI agents into a distinct managed category inside the operating system.
First, Apple split the bundled identity. The setting previously grouped under Device Control and Data Access is divided into two separate controls in macOS 27.3: Accessibility remains reserved for assistive tools, while cross-app control moves into a new setting called System & App Protection Override, signaling plainly that the app is bypassing built-in protections to control other software. Legitimate assistive and disk backup apps can apply for dedicated Apple entitlements to keep their smooth pre-27.3 onboarding flow, and currently known traditional apps remain temporarily unaffected. Other AI agents and automation tools go straight into the new override category and can no longer hide inside accessibility settings.
Second, Apple sliced all-or-nothing grants into fine-grained scopes. The new onboarding flow nudges users to narrow access: under Full Disk Access, you can approve just one or two specific project folders instead of the entire drive; under cross-app control, you can pick exactly which apps an agent may touch, keeping Signal or Mail off-limits. Apple also added an Ask Every Time option so users no longer have to hand out permanent passes, and explicitly told developers to make sure their apps gracefully handle access to only a limited set of folders or applications.
Third, Apple translated technical jargon into plain-language risk lists and added a live indicator for background control. According to beta screenshots reported by Macworld, when ChatGPT’s Mac client requests disk access, the new prompt spells out the consequences line by line: reading, changing, deleting, and exporting all data; viewing hidden and recently deleted photos; viewing browsing history and browsing as you. The menu bar gets a safeguard too: whenever an app uses cross-app override permissions to control another program, macOS displays a blue transparency indicator (an icon resembling a cursor over stacked windows) that lists which apps hold control. Much like the green camera light on a Mac, the beta offers no toggle in System Settings to hide it, even as early testers complained that ordinary window managers keep the blue icon lit.
The new rules do have clear limits. macOS 27.3 is expected to ship in December, and the changes apply only to new permission requests; broad grants approved before upgrading remain intact and must be tightened manually by the user. Apple is not banning high-privilege agents either—it is breaking hidden master keys into visible, narrower switches and handing informed choice back to the user.
Operating System Users No Longer Sit Only Outside the Screen
Put Apple’s overhaul next to the Reuters three-device test and a clear industry dividing line emerges. In the test, the iPhone stopped Muse from reading Signal thanks to strict mobile sandboxing, which isolates each app so one cannot touch another’s data by default. Android kept a notification-reading channel open, while desktop operating systems had to allow cross-app file access and window control for productivity. Those open interfaces were built assuming a human sat behind them; once AI agents took the wheel, desktop openness turned into a privacy funnel.
After the replication results surfaced, the debate began shifting from AI vendors to operating systems themselves. Signal president Meredith Whittaker criticized Meta’s rollout as technologically reckless and called for operating systems to provide a consistent technical way for privacy apps like Signal to mark themselves off-limits to AI agents, enforced by the OS rather than left to AI companies’ goodwill. Google, responding to Reuters, still emphasized that Android users must deliberately grant notification permissions, though it also acknowledged considering additional controls.
Looking back at Meta’s stance makes the contrast sharper. Mark Zuckerberg previously cited Muse’s isolated cloud virtual machines to dismiss calls for industry-wide AI safety standards, while Muse’s terms of service hold users responsible for the agent’s actions. Reality has shown that no matter how airtight a vendor’s cloud whitepaper looks, as long as an agent reaches onto your phone or laptop screen to get work done, the real safety boundary sits at the local operating system’s permission gate.
For forty years, operating systems were built solely for the human outside the screen: interfaces for human eyes, notifications for human reading, and permission toggles backed by human judgment. With macOS 27.3 carving out cross-app control, per-app limits, and a persistent menu bar indicator, operating systems have formally entered an era of human-agent cohabitation. They must serve the person at the keyboard while guarding against the AI agent sitting inside the machine, ready to click on your behalf. The next time you install an always-on AI assistant on your phone or computer, look past the vendor’s cloud-sandbox pitch and check your system settings first: is it asking only for specific folders and apps, or is it still reaching for the master key under the banner of Accessibility and Full Disk Access?