PyTorch v0.1.1 tensor metadata OOB disclosure PoC
Benign proof of concept for a valid exact PyTorch v0.1.1 tar artifact whose tensor metadata advertises 10 float32 elements while the backing storage only serializes 3.
Files
model/v011-oob-size10.pt— crafted exactPyTorch v0.1.1artifact (10240bytes, tensor size=10, backing storage=3)model/v011-clean.pt— clean baseline exactPyTorch v0.1.1artifact (10240bytes, tensor size=3, backing storage=3)build-meta.json— pinned hashes and replay metadatareproduce/replay_v011_oob.py— Dockerized exact-runtime replayreproduce/Dockerfile.v011— historical runtime build recipe from the upstreamv0.1.1taggenerator/build_v011_oob_artifacts.py— deterministic artifact generator
What it demonstrates
On the exact PyTorch v0.1.1 runtime, the clean artifact remains a 3-element tensor throughout normal operations. The crafted artifact loads as a 10-element tensor, and normal arithmetic paths like clone() and add(0) materialize stale tail values that were never present in the serialized storage.
Quick repro
Run python reproduce/replay_v011_oob.py. The script builds a CPU-only historical runtime from the upstream v0.1.1 tag in Docker and prints the clean-vs-crafted differential.
This PoC is for defensive validation and bug-bounty triage only.