
Research artifact for CVE-2026-112611, a CoreGraphics PDF ShadingType5 out-of-bounds write, with PoC generators and verified write primitives in the privileged CGPDFService.
Research artifact (v0.4). Verified PDF write primitives on the vulnerable build; no code execution yet. Reproduction reports, static analysis, and research on the remaining stages are welcome - see CONTRIBUTING.md, ROADMAP.md, and SECURITY.md.
Apple fixed an out-of-bounds write in the CoreGraphics PDF stack in iOS 27.2 / iPadOS 27.2; the fix is credited in the iOS 27.2 security advisory (HT149230) as CVE-2026-112611.
A PDF page can make CoreGraphics write to a memory address chosen by the document.
Tested on iPhone18,1 / iOS 27.0.1 (24A446) (the vulnerable build). The primitive was
observed in an ordinary app process, in a fault-guarded harness, and in the privileged
QuickLook PDF renderer CGPDFService (CGPDFService.xpc, reachable from
QLThumbnailGenerator / qlstart). A public fix-diff analysis of iOS 27.2 attributes the
defect to _cg_shading_type5_draw and identifies the 27.2 bounds-check gate; see
Credits.
Works - PDF-selected 4-byte zero write. A ShadingType 5 (lattice-form Gouraud) shading
whose /Function returns more components than the colour space has overflows the per-vertex
record. The draw code later executes:
df0e00b9 str wzr, [x22, #0xc]
with x22 derived from spilled fields, so the PDF chooses the address. The store is a fixed
4-byte zero (value is constrained by the instruction encoding), and it faults on read-only
and PROT_NONE pages - which is how the address control is proven (write fault, WnR=1,
far equals the chosen address).
Works - attacker-controlled heap spill with per-component values. With
VerticesPerRow >= 130 and a large /Function output count, the record buffer is allocated
on the heap and the per-vertex evaluation writes attacker-chosen IEEE-754 doubles past the
chunk. Each output component is controlled independently (component k lands at a
deterministic offset beyond the chunk). This turns the bug into a controlled heap
overflow, not just a crash oracle.
Works - arbitrary-target write of a heap pointer in CGPDFService. In the privileged
QuickLook PDF service, the spilled values are later reused by ColorSync's CMMMemMgr block
allocator: a spilled value becomes the next field of a CMMBase block, and
CMMBase::NewInternal performs str x19,[x10] - a store of a fresh CMM block pointer P
into the target address X if [X] == 0. Consequences on device:
X read-only: SIGBUS/KERN_PROTECTION_FAILURE, far = X - proves the store executed at
the chosen address;X mapped and writable: the write succeeds and the service survives and keeps serving
thumbnail requests (a live arbitrary-target write primitive in a privileged system
process);X non-zero: the list walk dereferences X, and the crash report's far reveals the
qword stored at X (a read oracle for machine-word checks).The spill index that becomes the block's next field was mapped exactly
(out[1231] = next for the 64 KiB page-rounded chunk layout used here), so a single-field,
low-noise document is available (out[1231] = TBI(X), all other outputs 0.0).
Does not work (yet). The primitive writes a fixed zero or a heap pointer, not an
arbitrary value. There is no arbitrary read, no data-only read/write primitive, and no
control-flow hijack at this point. On arm64e, all indirect calls in the graphics/color code
are pointer-authenticated (blraa/blraaz), so turning a data-corruption primitive into
execution requires further work (a write-through consumer, a signed-pointer replay path, or
a signing oracle). Code execution in CGPDFService remains the goal, not a current result.
What the v0.2 work added (details in the linked analyses and evidence):
mAB offset class is rejected; CMYK profiles are not applied through the tested
image path. The parser is hardened against this mutation style
(analysis/icc-parser.md).What the v0.3 work added:
Z + 0xcac (the same spilled value selects both), and
the whole pair fires once per cold ColorSync transform
(evidence/store-coupling.md).braaz/blraaz); no plain-blr hijack site exists.What the v0.4 work added: