0xiviel/poc-armnn-getnumelements-overflow

Arm NN GetNumElements() Integer Overflow (CWE-190)

Vulnerability

Root cause: TensorShape::GetNumElements() in src/armnn/Tensor.cpp:190-197 multiplies tensor dimensions using unsigned int without overflow checking.

When a crafted model provides large dimensions whose product overflows 32-bit unsigned, GetNumElements() returns 0. This propagates to GetNumBytes() (line 429), which is used for ALL buffer allocation and validation. The buffer size check at TfLiteParser.cpp:270-271 compares against the overflowed value (0 > N = false), so validation silently passes.

Affects ALL model formats: .tflite, .onnx, .armnn (native)

Vulnerable Code

// Tensor.cpp:190-197
unsigned int count = 1;
for (unsigned int i = 0; i < m_NumDimensions; ++i)
{
    if (m_DimensionsSpecificity[i])
    {
        count *= m_Dimensions[i];  // NO OVERFLOW CHECK
    }
}

Additional Finding

OnnxParser.cpp:2332 uses reserve() instead of resize(), then accesses elements by index (UB/OOB write).

Impact

• CWE-190 (Integer Overflow) + CWE-122 (Heap Buffer Overflow)
• CVSS 8.8 (Critical)
• Deployed on millions of embedded/mobile devices
• 30+ affected code paths across all parsers
• numeric_cast overflow checks disabled in release builds (NDEBUG)