Skip to content

About

Educational iOS 27 security research: how SPTM/TXM, the kernel's mitigations and the media parsers hold up, read from the shipping firmware, with a measurement bench. Negative results documented; no exploit published.

Topics

Resources

Code of conduct

Contributing

Security policy

Accessibility

Stars

62 stars

Watchers

1 watching

Forks

Bench27

Bench27

iOS 27 security research bench · iOS 27.0.1 (24A446) · iPhone 18,1 · A19 Pro

⚠️ Purpose & scope (please read first)

This repository is an educational security-research notebook. Its goal is to understand how modern iOS defends itself — the monitors (SPTM/TXM), the Secure-Kernel "exclave" world, the kernel's zone-pointer mitigations, and the media parsers — by reading the shipping firmware and writing down, clearly, how each mechanism works and where its trust boundaries are.

Everything here is static analysis of publicly downloadable firmware (Apple IPSWs) and, where a device is involved, of the author's own research devices with the owner's authorization. No third-party systems are touched.

No weaponized exploit, no 0-day, no jailbreak is published here. Where a vulnerability is discussed, it is either (a) already fixed by Apple in a later release — documented as a patch-diff, the way public security writeups do — or (b) a trust-boundary observation (for example: "mitigation X assumes the attacker does not yet have capability Y") — not a working attack. The recurring result across this research is, honestly, "hardened — no exploitable bug found." That negative result is itself the point: it documents why current iOS is hard.

The author does this kind of work for responsible disclosure (Apple Security Bounty) and for CV and skill development. Nothing in this repo is intended to enable harm to real users or real systems, and it must not be used that way.

The measurements are observations on one build, not guarantees, and the bench is a lab harness: some probes send malformed input to kernel drivers and can panic the kernel, restart the UI or reboot the device, so run them only on hardware you own.

This is not affiliated with or endorsed by Apple. "iOS", "Apple Silicon", SPTM, TXM, and the rest are Apple's.

If you are from Apple PSIRT and something here concerns you, please open an issue — happy to adjust or take it down.

Security research notes and a reproduction bench for iOS 27, run on a real device from a sandboxed app with no entitlements.

Property Value
Device iPhone 17 Pro · iPhone18,1 · A19 Pro (T8150)
OS iOS 27.0.1 (24A446), arm64e
Compared against iOS 27.2 (24B119) and 27.2.1 (24B156), statically
Bench Bench27 — a console: JavaScript over the network, no compiled probes, 0.3 MB
Verdicts STATUS.md — one page, one vocabulary
Still open research/OPEN-WORK.md — what is not finished and what each line needs next

Directions

# Direction Verdict CVE
01 IOSurface size validation demonstrated CVE-2026-112477 · IOSurface, a race fixed in 27.2
02 Metal GPU VA partially demonstrated —
03 mediaplaybackd / FigXPC demonstrated CVE-2026-112706 · reproduced: a live descriptor to a file the sandbox denies
04 AppleAVE2 partially demonstrated CVE-2026-112705 · and -112689, -112470, -112469, -112468, -112466 — six writes and a read in AVEVideoEncoder
05 ANE reachability not demonstrated —
06 AppleM2ScalerCSCDriver not demonstrated CVE-2026-43655 · the teardown UAF, fixed in 26.5
07 Container path disclosure demonstrated —
08 Image parser surfaces not demonstrated CVE-2026-20687 · fixed in 26.4, not in the 27.2 bulletin
09 External leads not demonstrated —
10 Sandbox escape to other containers demonstrated —
11 HID property writes demonstrated CVE-2026-112515 · IOHIDFamily, but an out-of-bounds read, not this write
12 AppleKeyStore interface partially demonstrated —
13 CoreGraphics PDF shading demonstrated CVE-2026-112611, CVE-2026-112610 · malicious PDF, out-of-bounds write
14 CoreGraphics CMap not demonstrated same CoreGraphics PDF pair as 13
15 CoreGraphics fix diff demonstrated same CoreGraphics PDF pair as 13
16 Brotli block-switch demonstrated CVE-2026-111497 · "Node2Shell", critical; the guard we found is its fix, and 27.0.1 lacks it
17 Brotli consumers demonstrated same brotli CVE as 16 — our work is the delivery, not the write

The CVE column follows two Apple bulletins — iOS 27.2 and iOS 27.2.1 — because together they are the releases our own diff is taken against. A CVE is listed where a bulletin names the same component, and the cell says whether it is the same defect or a neighbouring one:

  • 04 is the closest match anywhere in this repository. The bulletin lists six AVEVideoEncoder issues — CVE-2026-112705 and -112689 and -112470/-112469/-112468 as out-of-bounds writes, -112466 as an out-of-bounds read — and direction 04 reached two denial-of-service defects in that driver while the addressed write stayed out of reach. The third-party reproduction of -112705 names NumberOfTemporalLayers as the trigger and the Chollima team as the discoverers.

  • 13, 14, 15 are one family: CVE-2026-112611 and -112610, a malicious PDF causing an out-of-bounds write. The CG-SEC-187882049 and CG-SEC-187882136 tags recovered from our own 27.2 diff belong to that work; the bulletin does not split the two tags between the two CVEs, so no attempt is made here to say which is which.

  • 01 is CVE-2026-112477, described as a race fixed by improved locking — the same kext we measured, but not the size-validation defect direction 01 is about.

  • 03 is CVE-2026-112706, reproduced on 27.0.1: the bytestream service opened the media-library database with its own privileges and returned a live descriptor, which the app then read, while its own open() of the same path was denied. A forbidden path answers -12203 and a nonexistent one -17913, so existence is disclosed too (212). -112688 is a different CoreMedia defect and is not what direction 03 measured.

  • 11 is CVE-2026-112515, an out-of-bounds read in IOHIDFamily; our finding is a string value that aborts backboardd through a getter, which is not the same thing.

  • 06 and 08 are not in the 27.2 bulletin because they were fixed earlier — 26.5 and 26.4.

  • 16 and 17 are CVE-2026-111497: "Processing maliciously crafted input may lead to memory corruption", an out-of-bounds write fixed by improved input validation in 27.2.1, the release our brotli diff is taken against. Direction 16 found the guard that release added and showed the guarded state is reachable but terminal; direction 17 delivered a stream to that state over the network. Neither demonstrated the write the CVE describes.

Its other entries — AirPlay, Bluetooth, CoreUtils, Messages and rapportd — name components this repository did not measure; brotli is the one line in that bulletin that maps onto our work.

A dash means no entry in either bulletin names that component. It does not mean no CVE exists.

Builds: docs/ci.md — GitHub Actions on push.

Kernel read/write and a jailbreak were not demonstrated on this configuration. The escape in 10 crosses a container boundary in userland; it is not kernel code execution. Every verdict, and what would change it, is on STATUS.md.


Layout

app/                    Bench27: Swift + C, one direct swiftc/clang build, no Xcode project
  Sources/              the console; archived probes live in ../archive/probes/
  Resources/            sample media
research/               one folder per direction, each with a README and its evidence
  01-iosurface-size-validation/
  02-metal-gpu-va/
  03-mediaplaybackd-xpc/
  04-appleave2/
  05-ane-reachability/
  06-scaler/
  07-container-path-disclosure/
  08-image-parser-surfaces/
  09-closed-external-leads/
  10-sandbox-escape-mobilegestalt/
  12-patch-diff/
  13-physical-primitives/
  14-reachability/
  15-scaler-methods/
  16-hid-property-types/
  17-applekeystore/
  18-cg-pdf-shading/
  19-cg-cmap-recursion/
  20-cg-fix-diff/
  21-brotli-block-switch/
  22-brotli-path/
  archive/              early material, kept for history
docs/                   CI options and how the build is triggered
tools/                  build trigger, artifact download, signing, report sync

Scripts named as work/… in the notes are the local bench harness — the CI token, the signing material, the kernelcache extraction and the analysis helpers. They live outside this repository on purpose: this repository holds the sources, the measurements and the evidence, not the operator's keys. Where a report quotes a work/… script, that is what it means; the repository's own tools are the four files in tools/.

The evidence is committed as-is: session logs and crash reports under research/*/logs/ or next to the report that cites them, kept byte-exact (the .gitattributes in research/16-hid-property-types/logs/ turns off line-ending conversion for them).


Directions in brief

01 · IOSurface

A surface whose declared header region is 0x100000000 (4 GiB) is accepted while the real allocation is 64 KiB. The cause is 32-bit arithmetic on the compressed-plane geometry. No expansion of accessible memory was observed.

02 · Metal GPU VA

Access through a device pointer beyond the logical bounds of an MTLBuffer reached other GPU pages of the same context: reads and writes confirmed with readback. This is observed behaviour of the implementation, not a guarantee of the Metal API.

A two-process test (two apps, identical addresses) showed different physical pages, so access to another process's private buffers was not reproduced.

03 · mediaplaybackd

The FigXPC protocol was reconstructed end to end: handshake, asset creation from a URL and from raw bytes, property loading, frame decoding into an IOSurface. On that infrastructure, 120+ mutations of the parser and decoder produced no memory corruption.

Separately, a NULL dereference in sandboxserver was confirmed: a single message crashes the privileged daemon. A disclosure package is included.

04 · AppleAVE2

Two defects in the kext: a dereference of a pointer whose value is 1, and a fill loop with an externally supplied count. Both are reachable from an app through VideoToolbox, both are denial of service: the write address is fixed and the affected range is limited to its own mapping.

05 · ANE reachability

H11ANE and AppleT8150ANEHAL accept no client type, and no ANE daemon answers. One service does open — H1xANELoadBalancer, for client types 1 and 4 — and on both the only shape an unentitled client can call is selector 1 with no arguments. 660 calls across selectors 0–10, three output sizes, five input sizes and two scalar counts: everything else is BadArgument. Whether that one call does anything on the SEP side is not demonstrated.

06 · Scaler

AppleM2ScalerCSCDriver opens for all four client types from the sandbox. The ABI was mapped selector by selector, and one behaviour turned out to matter: the driver refuses every call until a context exists, and the context is created by ordinary scaler use — a single VTPixelTransferSession conversion turns NotReady into ok.

With that done, the transform accepts a real surface pair, and the property call accepts any payload. The read-back buffer, however, is static client configuration: fourteen probes, including ring fills and concurrency, produced zero bytes of change. The direction is closed and documented rather than left half-explored.

07 · Container paths

fsgetpath resolves an inode to a full path. Walking inode numbers from inside an ordinary app container disclosed 235 container paths the sandbox is not supposed to reveal. The paths are statable; listing them and opening files inside them is denied. It is disclosure, not an escape into data.

08 · Image parser surfaces

AppleJPEGDriver accepts a user client of every type without entitlements, which makes it a parser that receives bytes we control. Its selectors reject every structure shape tried so far — sizes up to 0x2000, zero to three scalars, all four client types, with and without a notification port. A 400-image mutation campaign decoded without a fault.

09 · Closed external leads

Three published directions were checked against this build and closed with evidence rather than assumption: the macOS AppleMultitouchDeviceUserClient UAF (the client cannot be opened here), the AppleAVE2 signed-count write (its entry point is external selector 5, and that client is not permitted), and AppleVideoToolboxParavirtualization (fixed in iOS 27).

10 · Sandbox escape

The container manager mints a sandbox extension for a path inside your own App Group. Replace that path with a symlink and the token lands on the target. On 27.0.1 the grant reaches other applications' data containers, app group containers and the shared system groups — read and write — and /var/containers/Data/System read-only. With write access to the gestalt cache, the About screen's Model Name and Model Number can be changed; writing a file in another app's sandbox was proved as well. Only class 7 answers, and renaming the app's own Library is refused, so the App Group entitlement cannot be avoided. The technique is worse_query by Taj C (forcequit); the measurement is ours.

Serial Number and the device name are not reachable this way: one is not in the cache, the other lives outside SystemGroup.

11 · HID property writes

A CFString written into ReportInterval aborts backboardd with an uncaught -[__NSCFString _getValue:forType:] inside IOHIDServiceGetReportIntervalForClient. The value is never parsed — "1000" behaves exactly like "bench27" — and of ten keys only ReportInterval takes the failing route; a numeric value is accepted and reads back as 65535. Two aborts seven seconds apart also leave SpringBoard stuck until a hard reboot. It is a denial of service in a system daemon, not memory corruption.

12 · AppleKeyStore interface

The user client opens from the sandbox and answers semantically once one to three scalars are present: 25 of 37 selectors return a real code (ok, NotPrivileged, NotAttached, NotOpen), where the earlier "three selectors, all zeros" was an artefact of calling with no scalars at all. What those selectors do with the scalars is not established; nothing here is a capability.

13 · CoreGraphics PDF shading

In a type-5 shading the renderer executes str wzr, [x22, #0xc] — a four-byte zero write at an address the document chooses, because the base of the store is document-supplied. The catch is what the value must be: only a normal double reaches the store, and a normal double names an address at or above 4.5 × 10¹⁵, which is unmapped. Denormal bit patterns can name a mappable address, but they never reach the store (195–197).

14 · CoreGraphics CMap

A genuine Identity-H stream from the Adobe resources is parsed, and codespaces of 4, 5, 6, 8, 16 and 64–4096 bytes are all accepted — the font is built and the document is not rejected. The copy in _csr_create that would matter is never reached: ISO 32000-1 §9.7.6.2 and Adobe TN #5014 limit codes to one to four bytes, and a wider stream is an invalid document, which is what 189 and 190 measured.

15 · CoreGraphics fix diff

The 27.2 build adds a check that the range dimension matches the colour space — the CG-SEC-187882049 and CG-SEC-187882136 tags recovered from the diff belong to it — and the public record for that work is the CoreGraphics PDF pair CVE-2026-112611 and -112610. Our own diff and the bulletin describe the same family; nothing here reproduces the write.

16 · Brotli block-switch

This is CVE-2026-111497, "Node2Shell", critical and in CISA KEV: the decoder ignored the block-switch status on its fast path, and the upstream fix returns -17 (ERROR_FORMAT_BLOCK_SWITCH) when fewer than two block types are declared. Our 24A446 and 24B119 libcompression copies discard the status; 24B156 checks it. The write is not reproduced here: the published trigger runs the distance block counter, while 167 walked the literal one, which reaches zero exactly as the meta-block ends.

17 · Brotli consumers

The delivery side is measured end to end: Content-Encoding: br decodes a 21-byte body (176), the server chooses the chunk boundaries and every split decodes (177), and the 13-byte stream from 167 returns 16 777 216 zero bytes over HTTP (178). So a network response reaches the decoder state in the process that handles it. Non-terminal use of that state is what is still missing.


Building

Two paths, both clang + swiftc against the iphoneos SDK and packaged into an IPA:

A commit carrying [skip ci] starts nothing.

  • GitHub Actions, workflow build.yml, runs only on manual dispatch, in five to fifteen minutes depending on the runner queue. It is the fallback when a bundle is needed without a new commit.

The artifact is signed locally and installed on the device. Details: docs/ci.md.

How to read the reports

Every claim is tagged with its source:

tag meaning
[I] confirmed by the source code
[L] confirmed by a runtime log from the device
[B] confirmed by static analysis of a binary
[G] hypothesis, needs verification

Reports are living documents: later sections supersede earlier ones, and each starts with a status table of current conclusions.

Tested configurations

device SoC OS result
iPhone 17 Pro (iPhone18,1) A19 Pro 27.0.1 (24A446) all findings measured here
iPhone 16 Pro Max (iPhone17,2) A18 Pro 27.2 (24B119) 01–07 reproduce (AGXAcceleratorG17P, same VA); the direction 04 temporal-layer trigger no longer panics — NumberOfTemporalLayers is rejected with -12900, which is CVE-2026-112705 fixed

Anything outside this table is untested. The bench prints a warning when the OS is not 27.0.1.

License

Documentation and original analysis © SKVLENT, released for educational use — see LICENSE. Firmware, symbol names, and Apple trademarks belong to Apple Inc.

The short version: read it, learn from it, quote it with attribution. Do not republish it as your own, and do not point it at systems you do not own. The measurements are observations on one build, not guarantees, and several probes are designed to crash or reboot the device they run on.

Accessibility

A statement of what this repository does for readers who use assistive technology, what it does not support, and how to report a barrier: ACCESSIBILITY.md.

Credits

This work builds on published research:

material author used in
F1OOB-PoC @murkuwuwu 01, 02
cinder27, ave27 @yang_niubi 03, 04
AppleResearch @dawgyg 08, 09
worse_query, bad_query Taj C — @forcequitOS 07
GestaltX huami — @huami_1214 10

Details and the exact corrections made to the original claims: CREDITS.md.

Limitations

Everything was measured on a single configuration: iPhone18,1, iOS 27.0.1 (24A446). Results must not be carried over to other SoCs or OS versions without verification. The absence of a finding here does not mean the absence of a vulnerability.

About

Educational iOS 27 security research: how SPTM/TXM, the kernel's mitigations and the media parsers hold up, read from the shipping firmware, with a measurement bench. Negative results documented; no exploit published.

Topics

Resources

Code of conduct

Contributing

Security policy

Accessibility

Stars

62 stars

Watchers

1 watching

Forks

Contributors

Languages