Apple posted macOS 27.3 beta 2 on November 16, 2026 with redesigned permission flows for Device Control, Full Disk Access, and Accessibility—the same macOS generation that already carries the Siri AI developer preview. The release does not rip away access from desktop AI tools or coding assistants, but it splits “control everything” from “help me with specific apps”, surfaces stronger warnings before you grant Mac-wide data access, and adds a menu-bar indicator when one app controls another. Previously granted permissions stay active until you revoke them manually.
If you run Claude Code, Codex, Cursor, or ChatGPT’s Mac app beside a terminal with real project folders, this changes what you should click “Allow” on—not whether agents can still exfiltrate data through paths you already approved.
| Question | Short answer |
|---|---|
| Am I affected? | Yes on Mac if any AI or automation app has Full Disk Access, Accessibility, or Screen Recording |
| Is there a fix? | No CVE—review System Settings → Privacy & Security and downgrade broad grants where you can |
| Does data leave my machine? | OS permissions are local; network exfil is unchanged—agents still send what you pipe to APIs |
| What is new vs older Apple privacy posts? | Split APIs, per-app pick lists, folder-scoped Full Disk Access nudges, ask-every-time option, control transparency chip |
| What should I do today? | Audit existing grants, prefer folder limits, keep coding agents out of Messages/Mail scope unless required |
explainx.ai summarized Siri AI daily limits on the same beta train; this post stays on permission mechanics from Apple’s developer announcement.
What Apple changed, step by step
Apple’s November 16 developer news item lists concrete UI and API policy shifts:
- Split Device Control and Data Access into two settings: System & App Protection Override and Accessibility. That replaces a single mental bucket where “accessibility” often meant “the app can drive my whole desktop.”
- One approval pattern: users pick the requesting app from a list instead of unrelated toggles scattered across panes.
- Scoped defaults: macOS guides users toward specific apps (for System & App Protection Override) or specific folders (for Full Disk Access) rather than whole-disk approval.
- Explicit risk language: approval screens must explain which categories of data full permission could reach—Apple states plainly that granting full access puts all Mac data at risk.
- Ask every time: users can configure apps to prompt on each access instead of a permanent grant.
- Transparency indicator: when an app controls another app via System & App Protection Override, macOS shows a menu-bar indicator (Apple’s November 16 post confirms the feature; beta testers reporting in Cyber Security News describe menu-bar placement).
Apple positions the work as empowering creators and developers while keeping users in control of privacy—language consistent with prior Full Disk Access tightening discourse, but this drop is the first that names AI assistants in downstream reporting because Siri AI’s desktop integration ships in the same beta generation.
Existing permissions are not reset
Apple’s text is explicit: “All app permissions previously granted by users will remain in effect.” Installing macOS 27.3 beta 2 does not narrow a ChatGPT, Claude, or automation tool that already holds Full Disk Access. Security value lands on the next approval dialog and on audits you run yourself in System Settings.
That matters for coding-agent users who clicked through prompts once during a hackathon and forgot: the OS will not save you retroactively.
Developer obligations
Apple tells developers to:
- Adjust onboarding when users choose limited folders or apps.
- Handle graceful degradation when access is partial.
- Plan for entitlements if they ship backup software (Full Disk Access) or assistive technology (Accessibility). Known apps keep prior behavior without entitlements for now; Apple promises more entitlement detail soon.
If you maintain an MCP bridge or local agent harness that relied on Accessibility to click other apps, expect support tickets from users who pick single-app scope.
How this intersects with coding agents on Mac
Coding agents touch Mac permissions in three overlapping ways:
- Dedicated desktop apps (ChatGPT, Claude, Copilot) that read files, capture screens, or automate UI.
- IDE extensions (VS Code, JetBrains) that request folder trust—a different gate, but often paired with terminal subprocesses that inherit your user account.
- CLI agents in Terminal (
claude,codex) that do not always trigger TCC for every read when you run them inside an already-trusted working tree—but can trigger Automation, Accessibility, or Full Disk Access when they drive GUI apps or scrape outside the workspace.
Apple’s split addresses category (1) and GUI automation paths in (3) more than raw shell reads inside ~/Projects. Narrowing Full Disk Access still helps when an agent’s companion app could otherwise read Mail, Messages, or Photos after prompt injection—similar to how browser-resident agents inherit the whole cookie jar of a logged-in session.
Secondary coverage (Cyber Security News) notes permission screens that mention messages, email, photos, and browsing for ChatGPT and a choice between overriding system protections vs selecting apps for Claude. Those are disclosure screens, not evidence those apps misused data—but they document how wide “full” still is if you click the wrong button.
For Claude Code’s VS Code extension, Apple’s OS controls are orthogonal to the workspace trust and symlink issues we covered in Claude Code VS Code symlink permission research: OS privacy gates Mac data classes; extension trust gates repository path surprises.
What this means if you run AI coding agents
Concrete checks on a Mac you use for agents:
- Open System Settings → Privacy & Security and sort by Full Disk Access, Accessibility, Automation, and Files and Folders. Remove anything you do not recognize or no longer use.
- Prefer folder-scoped grants when macOS 27.3 beta offers them. Point agents at a single client directory instead of your home folder.
- Do not grant System & App Protection Override to a chat app unless you need cross-app automation. For most coding workflows, terminal + repo-scoped tools suffice.
- Turn on ask-every-time for experimental assistants until you trust their tool policies.
- Watch the menu-bar transparency indicator after upgrading beta; if it lights during idle time, treat that as an incident triage signal, not noise.
- Separate roles: run production keys on a Mac profile without desktop AI Full Disk Access. Our lend remote-agent pattern is one way to keep gh and SSH off disposable hosts; Apple’s controls are another layer on the laptop you keep.
Commands and versions (from Apple, November 16, 2026):
- Beta OS: macOS 27.3 beta 2 announced on Apple Developer News.
- Paired tooling: Apple’s November 3 post lists Xcode 27.3 betas for the same SDK generation—rebuild Mac agents if you ship App Intents or Siri AI integrations.
Apple does not publish a shell command to list TCC databases; use System Settings UI or your MDM console. For headless CI Macs, keep agents off GUI permission classes entirely.
Relationship to Siri AI and third-party App Intents
The same beta wave opens Siri AI app intents to all developers (Apple’s November 3 announcement). That increases competition for “default assistant” surfaces—and multiplies Automation prompts when users ask Siri to drive your app while a coding agent also holds Accessibility.
Apple also documents daily usage limits on server-backed Intelligence features (including Siri AI) with future paid expansion—operational detail for product owners, not a permission bypass.
Teams shipping skills that shell out to Mac apps should read Apple’s promise of follow-up terms and usage policies for Siri AI experiences; Apple says App Store submission opens in spring, with customer shipping a few months later per the November 3 developer post—not a permission bypass today, but policy text may change before GA.
Threat model: what Apple still does not solve
Apple’s controls are consent and visibility, not content inspection:
- Prompt injection that tricks an already-trusted agent into exporting a repo still works inside approved folders—see skill and MCP supply-chain patterns.
- Malicious skills that run
curl | bashdo not need Full Disk Access—Beam’s heuristic scan catches known pipe patterns before execution, but OS privacy tiers do not replace command review. - Cross-device agents (remote VM + laptop credentials) bypass Mac TCC on the remote disk even when the laptop is locked down.
Use Apple’s update as a blast-radius reducer, not a substitute for tool allowlists, secret rotation, and local audit logs.
What Beam can and cannot do here
Beam watches Claude Code and Codex PreToolUse (and related hooks) on machines where you install the collector. It flags risky shell and MCP patterns, scans skills with eleven heuristics, and can deny when workspace policy is in enforce mode—but it does not manage macOS TCC databases or replace Apple’s permission UI.
Where overlap exists:
- Advisory mode helps you notice when an agent tries broad file reads after you thought you scoped a project—useful alongside revisiting Full Disk Access.
- Skill scanning catches “read entire home directory” tropes in third-party skills even before macOS prompts appear.
- Limits: Beam is hook-based, heuristic, and fail-open if cached policy expires; it is not a kernel sandbox and does not block GUI Automation grants you already approved in System Settings.
For local-first observability without shipping prompts to a cloud console, see how Beam introduced on-device agent security and the coding-assistant threat model primer.
Frequently asked questions
The FAQ front matter mirrors the questions above; short anchors:
- macOS 27.3 AI permissions — split overrides, stronger warnings, optional per-request prompts.
- Claude Code Mac permissions — separate IDE workspace trust from OS privacy panes; audit both.
- Is Full Disk Access safe for AI apps? — Apple warns it exposes all Mac data; prefer folder scope when offered.
Related reading
- Claude Code VS Code symlink and permission prompts
- MCP and skills: practical security guide
- lend: remote coding agents without copying SSH keys
- Introducing Beam: local-first agent security
- AI agent security for coding assistants
- Prompt injection in untrusted web content
Sources
- Apple Developer News — Device Control and Full Disk Access update (November 16, 2026)
- Apple Developer News — Siri AI developer preview (November 3, 2026)
- Cyber Security News — macOS 27.3 security controls (November 17, 2026)
Permission names, beta numbering, and developer obligations are accurate as of November 17, 2026, per Apple’s developer news posts. Beta behavior can change before general availability.

