Flatbuffer Size Metadata OOB Read
Defensive validation / triage only.
This PoC shows that a crafted flatbuffer-exported JIT artifact can widen a
tensor's logical shape through serialized TensorMetadata while keeping the
backing storage at one float. The loader accepts the artifact, and normal
inference then reads a second logical element outside the serialized storage
extent from artifact-controlled padding bytes.
Files:
model/clean.ff: clean controlmodel/evil_size2.ff: crafted flatbuffer artifact with widened logical tensor viewmodel/build-meta.json: hashes and generation metadatareproduce/replay_flatbuffer_size_stride_oob.py: replay script for clean vs crafted load/inference
Quick repro:
python reproduce/replay_flatbuffer_size_stride_oob.py \
--clean model/clean.ff \
--crafted model/evil_size2.ff
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support