PyTorch inductor op builds out-of-bounds CPU tensor views
Unchecked offsets in _reinterpret_tensor let callers produce views that read past storage and trigger heap buffer overflows.
torch.ops.inductor._reinterpret_tensor can return a tensor view that does not fit inside its source storage, so a later read on CPU becomes a heap buffer overflow.
The operator takes size, stride, and offset metadata and builds a new view without verifying that the resulting layout stays inside the base tensor’s allocation. A minimal case is enough: start from an eight-element int64 tensor, ask for a one-element view at offset 8, and calling .item() on the result reads eight bytes past the end of the 64-byte block. AddressSanitizer builds flag the fault on the dense scalar path.
That matters because _reinterpret_tensor is part of the inductor surface used to reshape and alias storage during compilation and runtime. Callers that trust the returned view as in-bounds can hit memory corruption on ordinary CPU execution, not only under specialized backends. The issue was reported against pytorch/pytorch by ydwu4 with a short reproducer and ASAN trace; no CVE identifier is attached in the report.