Status: local prototype controls implemented; internet desktop and E2EE NOT validated or release-ready.
The Mac host OS, account and installed native/control binaries are trusted. A compromised host can read all Guest data. Guests are untrusted and may return malicious/oversized frames. Local MCP stdio callers receive only scopes configured by their launcher, but the same OS account controls daemon files and can launch a privileged client; local scopes are delegation, not a sandbox against the account owner. The Guest vsock service accepts only host CID2; it does not expose a guest network port. Guest outputs are untrusted data for an external Agent.
Paired remote devices are trusted only for explicitly granted VM scopes. Discovery, signaling, STUN, TURN and authentication are separate roles. Managed discovery/signaling/TURN operators are not trusted with desktop, clipboard, files, commands, or pairing replacement. They may observe connection metadata and deny service. Browser application code is in the trusted computing base: if hosted mutable JavaScript is compromised it can steal plaintext/keys. E2EE cannot protect against malicious code executing at either endpoint. Prefer a host-served pinned client or auditable immutable version with independently authenticated distribution; ordinary SRI from the same compromised HTML origin does not solve this.
- Implemented local pairing CLI and browser helper: independently compare the complete SHA256 pairing digest binding both P-256 keys, VM, scopes and random nonce. QR/short-code UX is not implemented. No automatic TOFU. Host private key is non-extractable in login Keychain; browser private CryptoKey is non-extractable and structured-cloned into IndexedDB. Actual browser refresh preserves identity, and explicit local fixture unpair invalidates the old signing handle. Same-origin script can still invoke its key.
- Implemented signed, versioned per-device/per-VM grants in a private durable registry. The loopback media integration permits observe, or explicitly approved input+observe for a configured QEMU controller. It reloads the registry every250ms and before admission/confirmation; real macOS browser revocation stops capture, reaps the hardware encoder and rejects reconnect. Browser input is bound to its paired session, the input scope and a fresh screenshot; MCP bearer grants remain a separate authorization system. Re-pairing after revoke increments the grant version; explicit browser unpair generates a new key on next creation.
- Implemented versioned canonical signed offer intents and host answers binding both identity IDs, fresh60s challenge/host-instance nonce, VM/scopes/grant version, complete UTF-8 SDP hash and SHA256 DTLS fingerprints. Signature verification and challenge consumption are atomic before peer reservation. Browser verifies before setRemoteDescription; a signed transcript confirmation gates native encoder launch. Signed close authorizes only teardown and retains a bounded240s proof for disconnect/revocation races; only confirmed cleanup returns success. Real localhost HEVC/H264 received frames after this handshake. Adversarial relay/Internet tests and a full security review remain outstanding.
- Desktop media uses peer DTLS-SRTP with authenticated fingerprint binding. Browser input and screenshot bodies now use a bounded reliable DataChannel on that authenticated DTLS association, with independent scope and frame checks. Bounded file transfer now uses the same authenticated channel with an independent files grant, session-owned transfer tokens and a private Guest directory; actual roundtrip/cancel/cleanup evidence is in browser files (
vm-archive/docs-evidence/browser-files-live/). Remote clipboard transport remains unimplemented. The local viewers (the browser desktop on this Mac and the native console) share a Linux desktop's text clipboard with the person viewing it; nothing reaches MCP agents through the daemon. Actual loopback browser/Guest tests pass; Internet/hostile relay verification remains outstanding. TURN sees encrypted traffic, addresses, volume and timing. - Direct remote MCP HTTPS terminates on the Mac, with principal/scopes and session lifetime enforced there. This is encrypted device-to-Mac transport, not a claim about a hosted reverse proxy. A TLS-terminating hosted reverse proxy would expose plaintext and is prohibited for sensitive bodies. Future hosted MCP forwarding requires a cryptographically authenticated application tunnel all the way to the Mac, plus a compatible Agent-side transport; do not call ordinary proxied TLS end-to-end.
- STUN/ICE credentials are short-lived and separate from control grants. Cloudflare long-term TURN key remains server-side. Mint narrow TTL credentials after authentication, rate-limit by principal/host, cap sessions/bandwidth and refresh carefully; an ICE credential is not VM authorization. Force-TURN tests must verify relay candidate selection and fingerprint tampering rejection.
Bounded native/daemon/guest messages; restrictive state files/socket; no shell interpolation in command runner; opt-in execute; bounded guest output and timeout; input only with the input scope and a fresh matching screenshot; explicit destructive force-stop confirmation. Unsupported screenshot/input/accessibility return errors or are not advertised. Installed Linux commands now run as fixed UID/GID1000 with no supplementary groups; real Guest tests reject root-path writes and direct poweroff while independently scoped lifecycle poweroff succeeds. Async jobs are keyed by VM and job id, run under a Guest-side timeout and are killed with their process group on timeout, cancellation or the end of their MCP session. Synchronous execution uses the same registered job path; actual installed Linux tests verify stdio/Unix disconnect, HTTPS DELETE and expiry cancel in-flight commands. Guest services lacking async job support fail explicitly rather than falling back to uncancellable execution. Production Linux execution uses root-owned per-job cgroup v2 subtrees with atomic clone3 placement and cgroup.kill; actual setsid/background-grandchild tests pass for cancellation, timeout, parent completion and MCP stdio EOF. Unsupported containment fails startup. The explicit legacy RAM diagnostic uses process groups, reports descendant containment false, and still permits setsid escape. Cancellation cannot undo completed effects or replace Guest OS privilege policy. These are tested narrowly, not a full security audit.
Linux service crash boundary: SIGKILL cannot run cleanup. A real test observed a still-populated job cgroup after the service died; the next exclusive service startup reaped its owned stale groups before accepting new RPCs. Both command and separate files-worker paths then recovered. Descendants may execute during the death-to-restart interval; no immediate crash cancellation or rollback of intervening effects is claimed. Cleanup errors stop new command admission, cancel peer jobs and prevent a false service-stop success report. Root-owned0700 cgroup directories prevent the fixed non-root command account from migrating itself out; the service and files worker are outside job groups. This is process-lifetime containment, not filesystem/network/syscall isolation, per-job resource limits, protection from Guest root compromise, or isolation between programs sharing an execution UID. See actual cgroup evidence (vm-archive/docs-evidence/linux-cgroup-2026-10-08/).
Still required: production pairing UX and signing/upgrade key lifecycle, all-media/control cryptographic verification, deployment of verified descendant containment across supported Guest systems, confirmed termination when Guest transport fails, complete least-privilege desktop Guest agents, identity-bound remote grants and browser-code distribution, arbitrary-path file access and upload-to-execution authorization, clipboard isolation, upgrade rollback, dependency/license audit, crash-dump and log policy. Do not expose development listeners publicly.
Diagnostics: committed evidence contains only deliberately generated fixture output and non-identifying host capability facts. Arbitrary user commands, screenshots, tokens and file bodies must not enter routine logs. Serial logs are opt-in fixture artifacts; default production config should omit them. No cloud storage or cloud resources have been created.
References: MCP transport, MCP authorization, Cloudflare short-lived TURN credentials, WebRTC security architecture RFC8827.
Linux files have an independent opt-in permission and dedicated fixed UID/GID1001 worker, separate from command UID1000. Only bounded single filenames under a private anchored directory are supported. Symlinks, hardlinks, wrong-owner files and special files are rejected; upload publication is no-overwrite and download completion verifies actual transmitted bytes. Session cleanup and independent Guest expiry bound unfinished transfers. This is not arbitrary Guest filesystem access; see evidence/control-files.md.
macOS files use an explicitly enabled separate worker with the same non-root UID
as the desktop integration service. Private directory anchoring, no-overwrite
publication, hash/size/offset checks and owner-bound tokens restrict remote
callers; they do not isolate files from other programs running as that UID.
Actual MCP EOF/idle cleanup and worker-death/restart tests pass. Worker death
fences files; reserved staging is removed on verified service restart, not
immediately at SIGKILL. Controlled host APFS ENOSPC preserves published data,
but Guest full-disk/power-loss persistence remains unverified. See
file evidence (vm-archive/docs-evidence/macos-files-live/). macOS command execution
stays disabled on the accepted path; a process group or launchd job does not
provide recursive cancellation after setsid. See
containment assessment.
The unsigned diagnostic remains available only when authentication flags are absent. Authenticated mode rejects legacy unsigned offer/close routes. Host/Origin checks still apply. This loopback-only implementation is not an Internet endpoint. Keychain ACL tests show different unsigned executables cannot share signing access: use the same frozen media executable for --pair administration and serving. Rebuild/update identity continuity requires a stable signed app and explicit lifecycle validation; do not change Keychain ACLs or fall back to exportable keys. See signed media evidence (vm-archive/docs-evidence/session-auth/) and protocol decision.
Signed browser input now requires a confirmed transcript, current paired grant, matching running QEMU incarnation/console inode, consumed monotonic sequence,5second request expiry, and an explicitly observed frame no older than5seconds. One in-flight action and no queued actions limit stale input. Expiry and revocation deny further requests. Session teardown cancels the in-flight request; the daemon releases held input at the end of each operation, which is not proof that an arbitrary Guest application consumed input. Actual input/revocation evidence is in evidence/browser-input-live/README.md. No plaintext loopback diagnostic listener may be exposed publicly.
The prior plaintext /auth/control endpoint has been removed. Paired browser controls require the same DTLS peer whose complete SDP/fingerprint is signed; DCEP channel naming alone grants no authority. Strict binary framing,1MiB logical message cap,16KiB fragments, one operation, bounded buffers/deadlines, monotonic signed sequence and short expiry remain independent checks. Malformed transport or denied authority fences the channel/session; there is no HTTP fallback. Cleanup recognizes only the RTC library's typed already-removed-channel state, while session cleanup and final peer closure remain required. See evidence/browser-control-channel/README.md for actual browser acceptance, original failure and negative cleanup tests. Signed signaling is still loopback HTTP and exposes metadata; this is not a complete Internet/E2EE release claim.
Browser negotiation cleanup distinguishes local transmission evidence: before invoking /auth/offer, a failed challenge/signature/aborted request cannot have admitted a peer, so local closure suffices. Once an offer is invoked, unsigned HTTP errors or missing/tampered answers cannot prove non-admission. A pinned-host signed exact-intent not_admitted rejection can establish non-admission, followed by confirmed local cleanup. For admitted sessions, the current loopback path requires a close response to the transcript-bound signed request; the response itself is not signed and does not establish Internet transport authenticity. Unknown cleanup remains fenced, including through unpair/dispose. See signed refusal evidence (vm-archive/docs-evidence/browser-offer-rejection-live/) and recovery evidence (vm-archive/docs-evidence/browser-cleanup-recovery-live/).
macOS console identity is not lock-state authorization. A real Guest Lock Screen
operation left the public identity eligible; title-free window enumeration and
screenshot still succeeded, while lock_state remained explicitly unknown.
Input was unavailable because AX is not granted; this is not evidence of a lock
fence when AX is granted. Observation does not promise suspension during lock,
and locking must not substitute for explicit grant revocation. Coordinate
anchors never authorize input by themselves. See the
actual lock-screen diagnostic (vm-archive/docs-evidence/macos-windows-lock-live/).