#90

1.6 and 1.6.1 pre-release confirmed to not work on watchOS26.6.1

Openprmntropened 6d ago

App fails to capture baseline mobilegestalt

"the MCM query returned no sandbox token - send Documents/bqtool.log"

See attached images

Happy to do more testing to get this working

iPhone 12,1 (MWLL2VC/A) iOS 27.0.1 (24A446)

Activity

10
sharryboy886d ago

I got an Apple Watch Ultra 1 with 26.6 (23U67) installed and managed to sideload the ipa with WatchBridge Windows, when i try to create a backup i get the error like you: MobileGestalt Access Denied Maybe a Group-ID or this watchOS-Version isn't supported. Check the App Group-Configuration and the watchOS-Version for this App.

erentomurcuk6d ago

Watch SE 2 running watchOS 26.6.1, same error - built from main and installed directly to phone and watch via Xcode.

Log findings (bqtool.log attached but sanitised from personal ids)

156 escape attempts Every try returns container metadata but NO sandbox token:

  • class7-symlink and class7-symlink-default: result non-nil, path resolves to the app group Library/Caches, token=NO consume=-5
  • class2-stage (all 3 variants) and class2-subpath (all 3 variants): token=NO consume=-5 (missing variants fail with posix=2 before the query)
  • class13-direct: result non-nil, path resolves to /private/var/containers/Shared/SystemGroup/systemgroup.com.apple.mobilegestaltcache/Library/Caches, token=NO consume=-5
  • class12-geod-traversal: result=0x0, token=NO

So MCM returns container objects including the MobileGestalt system-group container itself but container_copy_sandbox_token returns NULL for every class on this build.

The exploit fails at the token step, before ever reaching MobileGestalt.

bqtool.log

mustafaindap6d ago

On Apple Watch 26.6.1 new update is not working.

SebastienBreton6d ago

Another data point: Apple Watch Series 7 41mm (Watch6,8), watchOS 26.6.1 (23U95). Built from main @ 7d84d97 (1.6 pre-release) with Xcode, paid developer account, custom App Group present in entitlements.

Results

  • Backup MG: ✅ works, but only with a local read-first change (reading com.apple.MobileGestalt.plist directly, without escape). Reading is allowed on this build.
  • Enabling Hermès: ❌ "Unable to change setting".
  • bad_query (class13, 8-dot traversal, flags 0x8000000000) and class13-traversal: ❌ failed. The traversal step is disabled afterwards (gTraversalState = -1), so it ran and did not get a token.
  • class7 symlink with flags 0x800000000, 0x8000000000 and 0x8100000000: ❌ result non-nil, token=NO.
  • class2 stage/subpath, class13-direct, class12 geod: ❌ token=NO consume=-5.
  • Direct write probe on the MG cache: O_WRONLY and O_RDWR → Operation not permitted (errno 1).

Same picture as the 23U95 analysis in #7: reading works, every escape route returns no usable token, and writing is denied by the sandbox. So the experimental watchOS 26 support does not work on Series 7 / 23U95 either.

prmntr6d ago

1.6.1 confirmed to not work; same issues as above

Yatinmanjunath86d ago

Do we need a paid dev account to do this? Or a free dev account is enough? I got it installed on watch os 26.5 through Xcode and the installation was successful but same mobile Gestalt access failed message came up.

iarata5d ago

Not working on 26.6 Series 6. Same issue as MobileGestalt access failed.

MDX-Tom5d ago

Tested on S7, 26.6.1, hold the exact same opinion with @SebastienBreton:

Measured on a physical Apple Watch Series 7 (Watch6,6), watchOS 26.6.1 (23U95, arm64_32), with a build that repeats the upstream queries verbatim.

Direct access (no escape needed): reading com.apple.MobileGestalt.plist is allowed — 13202 bytes, CacheUUID 3BE19A76-F253-4CE4-AAC1-DF6D801CAAB1, CacheData 6309 bytes, 70 CacheExtra keys, matching the 23U95 entry in MGOffsets.plist. Writing it, or creating a file in the cache directory, fails with EPERM.

bad_query route (class 13 + part-domain traversal): a token is issued, but only with the low-bit flags 0x800000000 / 0x810000000 / 0x900000000 and set_group_identifiers; 0x8000000000 (bad_query's flag) returns no result. The token is read-only — the cache write-open and the directory create both still fail with EPERM.

worse_query route (class 7 + part 3 + symlink): no token at all for any of 0x8100000000, 0x8000000000, 0x0800000000, 0x800000000, 0x810000000, 0x900000000. Same for class 7 + part-domain traversal and for replacing the part directory with a symlink.

Control on iOS 27.2 with the same code: class 7 + part 3 + 0x8000000000 / 0x8100000000 issues a token and the cache write is allowed — the accepted flag family is the inverse of watchOS 26.6.1's, where class 13 is the only route still issuing a token and it is read-only.

API note: the extension handle is not a directory descriptor — F_GETPATH returns /dev/ttys000, so the target path must be used directly after sandbox_extension_consume.

So on 26.6.1 reading and backup work, but no route yields a writable extension, so the face and capability switches cannot be applied. Documenting 26.6.1 as read-only rather than unsupported may be more accurate.

kharonbridge5d ago

Additional physical-device results: Apple Watch Series 7 (Watch6,7), watchOS 26.6.1 (23U95).

These tests used a local diagnostic patch based on upstream 1.6.1, commit a4feb9071c2475cb25b54953ec60b368c054cd26, installed through Xcode with developer signing and Developer Mode enabled. This is not a test of the newer 1.6.2 release.

Backup and direct access

  • Reading the existing MobileGestalt cache directly, before requiring an MCM escape, succeeds.
  • The initial backup is 13,202 bytes. The exported backup passes plist validation and its SHA-256 matches the backup transfer manifest. It contains 70 CacheExtra keys and 6,309 bytes of CacheData.
  • A separate direct write-access probe returns EPERM / errno 1, using O_RDWR | O_CLOEXEC | O_NOFOLLOW, without O_CREAT or O_TRUNC.
  • This independently supports separating backup reads from write diagnostics, as proposed in #89. We tested our local implementation, not the exact PR branch.

Bounded exact-file query experiment

We also tested an exact-file variant of the existing class-13 / part-domain approach, following the public escapeBadQueryFile approach in worstgestalt. The query targeted the existing plist rather than its directory: class 13, group identifiers, part 3, eight parent components followed by the existing file path.

Four trials covered the /var and /private/var path spellings and two flag values:

FlagsResult across both path spellings
0x8000000000No token obtained; write-open not attempted
0x0800000000 (logged as 0x800000000)Token obtained and consumed, but write-open still returned EPERM

Both consumed tokens were released successfully (release=0). The original file content was compared byte for byte after every trial and remained unchanged. This probe performed no write, truncate, create, rename, symlink swap, or cache rebuild.

Selected anonymized log excerpts:

MG direct read succeeded: 13202 bytes (no system write)
MG initial backup created: initial (13202 bytes)
MG loaded: 70 CacheExtra key(s), 6309 byte(s)
MG write-access probe: denied errno=1 (Operation not permitted); no bytes written
FILE-PROBE v1 BEGIN model=Watch6,7 build=23U95 bytes=13202; no write/truncate/create/rename/rebuild
FILE-PROBE trial=1 original-content=UNCHANGED
FILE-PROBE trial=2 original-content=UNCHANGED
FILE-PROBE trial=3 original-content=UNCHANGED
FILE-PROBE trial=4 original-content=UNCHANGED
FILE-PROBE END trials=4 consumed-tokens=2 writable-descriptors=0 content=UNCHANGED

The measured result on this device is: backup works with the read-first change; these probes did not obtain write access, and native Hermès activation has not been achieved. Consuming a token alone is not evidence of write permission.

Only selected diagnostic lines are shared here. Device/account identifiers, personal paths, container UUIDs, memory addresses, and the cache/backup contents are omitted.

gfanour5d ago

I also tried version 1.6.2 on watchOS 26.6.1 and it doesn't work. Any ideas? Will a fix be released?