Within 24 hours of each other, OpenAI and Meta expanded their Mac apps with exactly the kind of features privacy advocates have warned about for years: access to private messages and to whatever is currently on screen. On August 19, 2026, Meta released a fundamentally rebuilt Mac app for Meta AI that reads along inside windows a user attaches to it. A day later, OpenAI followed with a ChatGPT update that lets the assistant read, summarize, and reply to iMessages, SMS, and RCS messages on its own.
Key takeaways
- ChatGPT on Apple Silicon Macs can now read, summarize, and reply to iMessages, SMS, and RCS messages on a user’s behalf, available across every subscription tier including the free one.
- The integration relies on AppleScript and macOS Accessibility features, which requires granting Full Disk Access plus access to Contacts and Automation tools.
- Meta AI arrived a day earlier as a standalone, natively built Mac app: users attach a window to the conversation, and the assistant reads its visible content to answer questions, without controlling other apps itself.
- Both companies say messages need manual confirmation before sending and that no fully searchable database of all messages gets built.
- The ChatGPT integration reaches deep into Apple’s own Messages ecosystem without Apple’s involvement, which observers flag as a likely friction point.
What ChatGPT can now do on the Mac
The new feature ships as an Apple Messages plugin and only runs inside the ChatGPT Work and Codex modes of the Mac app, not in the regular chat window. Technically, ChatGPT doesn’t use an official Apple interface for this; instead, it relies on AppleScript combined with macOS Accessibility settings, according to multiple outlets that tested the rollout. To make that work, the app needs Full Disk Access, Contacts access, and Automation permissions granted, three of the most far-reaching permission categories macOS offers. In exchange, ChatGPT can search conversations, catch a user up on missed threads, and draft replies. OpenAI built in one safeguard: by default, every reply needs individual confirmation before it actually gets sent. In practice, that means anyone who grants blanket approval without thinking hands an AI system direct control over their communication with family, coworkers, and government agencies.
The feature is limited to Apple Silicon Macs; older Intel machines are excluded entirely. OpenAI also stresses that processing happens locally on the device and that no fully searchable index of all messages gets created, a detail meant to offset the scope of the permissions involved. How reliably that promise holds up in practice is hard to verify from the outside, since OpenAI hasn’t disclosed the technical implementation.
Meta AI: the screen as context
Meta’s approach differs on one key point: rather than running quietly in the background, users have to actively attach a window to the conversation before Meta AI can see it. The newly rebuilt app, written natively in AppKit and SwiftUI, then reads both the visible text and a screenshot of that window to answer questions, suggest edits, or explain content. That requires the macOS permissions for Screen Recording and Accessibility. Despite the name, the feature stops short of letting the assistant click or type inside other apps itself, something competitors already offer under the label computer use. Meta describes its own approach as context, not control: the app answers based on what it sees but doesn’t actively act inside other applications. The release also adds system-wide dictation for inserting spoken text into any app, plus an archive of past requests. The app remains in beta, version 1.0, and requires macOS 15 or later on an Apple silicon Mac.
Two paths, one problem: how much access is too much
Both companies are solving the same problem, pulling the assistant out of its chat window and closer to users’ daily routines, with different trade-offs. OpenAI opts for deep, persistent system access paired with a per-message confirmation step; Meta opts for a deliberate, one-time invitation per window. For users, that produces two very different trust models: with ChatGPT, you grant the permission once and then rely on the send-confirmation prompt every time after; with Meta AI, you decide anew with every use which window the assistant actually gets to see. As our look at a tool exposing what ChatGPT really knows about its users shows, every new system integration grows the quiet data footprint companies build up, regardless of what the official privacy policy says. The fact that both companies moved almost simultaneously fits a pattern we’ve also traced in our piece on how people actually use AI day to day: assistants are increasingly expected not just to answer, but to act.
What this means for users
Anyone trying these new features should first get familiar with macOS’s Privacy & Security settings and check exactly which app has been granted which permission. With ChatGPT, it’s worth a second look at the Automation permissions specifically, since Full Disk Access goes far beyond what’s technically needed for reading messages alone. With Meta AI, the attach-a-window design limits the risk somewhat, but it still raises the question of whether every window, including ones showing banking details or health information, should really be handed over for a quick AI answer. It’s also notable that both companies are following a path similar to Google’s persistent agent, which we covered recently in our piece on Google’s Gemini Spark: the next competitive edge apparently isn’t just a better language model, but whoever gets the most direct line into users’ digital routines. Whether Apple keeps quietly watching this reach into its own Messages ecosystem is the real question to follow in the coming weeks.