Heap buffer overflow in cuJSONLinesIterator::gotoArrayIndex

Environment

• cuJSON commit: 2ac7d3dcd7ad1ff64ebdb14022bf94c59b3b4953 (branch master)
• OS: Ubuntu 22.04.5 LTS
• GPU: NVIDIA A40 (Ampere, sm_86)
• NVIDIA driver: 590.48.01
• CUDA toolkit: 13.1 (nvcc V13.1.115)

Component

cuJSON — JSON Lines iterator, cuJSONLinesIterator::gotoArrayIndex (paper_reproduced/src/query/query_iterator.cpp:269-315)

Severity

High (CWE-122, Heap-based Buffer Overflow).

Description

cuJSONLinesIterator::gotoArrayIndex mirrors the standard iterator traversal logic and can advance nextNode beyond the valid parsed structure when an operation sequence requests an invalid array element, producing a host-side heap-buffer-overflow.

Root cause

The JSON Lines iterator shares the standard iterator's traversal logic and likewise lacks bounds checks around nextNode, node, pair_pos, and structural-array accesses. An out-of-range array index drives the iterator past the parsed structure and reads past the heap buffer.

Call path:

harness LLVMFuzzerTestOneInput
  -> cuJSONLinesIterator
  -> cuJSONLinesIterator::gotoArrayIndex   (heap-buffer-overflow)

Impact

Malformed JSON Lines input plus iterator operations can crash the host process and risk adjacent heap access in services that expose iterator navigation.

Reproduction

CUDA_HOME=${CUDA_HOME:-/usr/local/cuda}
nvcc -std=c++17 -arch=native -O2 \
  -I<cujson-src> -I${CUDA_HOME}/include \
  -Xcompiler -fsanitize=address,-fno-omit-frame-pointer \
  llm_harness_cuJSONLinesIterator.cu harness_afl_main.cpp \
  --compiler-bindir g++-13 -L${CUDA_HOME}/lib64 -lcudart \
  -Xlinker -fsanitize=address \
  -o /tmp/cujson_cuJSONLinesIterator_asan.bin

ASAN_OPTIONS=protect_shadow_gap=0:detect_leaks=0 \
  /tmp/cujson_cuJSONLinesIterator_asan.bin trigger.bin

Sanitizer evidence

==3418215==ERROR: AddressSanitizer: heap-buffer-overflow on address 0x7b9d4b70457f
SUMMARY: AddressSanitizer: heap-buffer-overflow in cuJSONLinesIterator::gotoArrayIndex(int)

Suggested fix

Add explicit bounds checks around nextNode, node, pair_pos, and structural-array accesses. Reject out-of-range array indices before traversal mutates iterator state.

Attachments

The following files from this finding are attached:

• llm_harness_cuJSONLinesIterator.cu
• harness_afl_main.cpp
• trigger.bin
• asan.log

Archive.zip