README.md

metadata

license: mit
tags:
  - security
  - poc
  - executorch
  - webgpu

Harmless PoC: ExecuTorch WebGPUDelegateHeader unbounded parse

WebGPUDelegateHeader::parse(const void* data) reads a fixed header through byte 29 without a length argument. WebGPUBackend::init() calls it with only processed->data(), ignoring processed->size().

A 1-byte BackendDelegate processed buffer triggers AddressSanitizer stack-buffer-overflow at WebGPUDelegateHeader.cpp:78 (memcmp of magic at offset 4).

Files

• short_delegate_blob.bin — 1-byte processed buffer
• poc_parse_asan.cpp — standalone ASan harness calling parse()
• asan.log — captured ASan report @ executorch 8077bb60

Reproduce

# From pytorch/executorch checkout:
clang++ -std=c++17 -fsanitize=address -fno-omit-frame-pointer -g \
  -Isrc -I. \
  backends/webgpu/runtime/WebGPUDelegateHeader.cpp \
  poc_parse_asan.cpp -o header_asan
ASAN_OPTIONS=detect_leaks=0:halt_on_error=1:abort_on_error=1 ./header_asan

Harmless crash PoC only. Requires WebGPU backend sources (parse unit is independent of GPU hardware).