PointCloud Workflows is in development. Security fixes target the current main branch. This repository does not ship the original research source or its historical dependency environments.

Use Security → Report a vulnerability on the repository security page to send a private report to the maintainer. Include the affected revision, the smallest synthetic reproducer, expected and actual behavior, and the impact. Do not include access tokens, personal data, or a private research dataset.

If GitHub's private reporting form is unavailable, open an issue titled Security contact request containing only a request for a private reporting channel. Do not put the vulnerability or proof of concept in that issue.

Dataset and model readers validate their documented contracts. Numeric arrays are loaded without pickle, and supported weights and checkpoints use safetensors with JSON metadata. File hashes detect changes relative to recorded metadata; they do not authenticate an unknown artifact publisher.

Run unfamiliar inputs in an isolated environment with appropriate memory, disk, and execution limits. Do not treat Python package installation, legacy research scripts, or arbitrary dataset conversion code as a sandbox. The supported CLI and artifact contracts are described in the documentation.

Training run ownership uses local POSIX process locks. Shared network storage does not turn those locks into a distributed coordination service.

The default ml extra and CPU container use PyTorch 2.13.0, the first patched release for GHSA-rrmf-rvhw-rf47. The separate WarpConvNet GPU profile retains PyTorch 2.10.0+cu128 because its compiled extension requires that PyTorch/CUDA combination. That profile remains affected by the dependency advisory; a CPU upgrade does not resolve its alert.

The reviewed issue concerns torch.jit.script. PyPI also reports PYSEC-2026-139 for PyTorch 2.10.0, concerning untrusted .pt2 loading through torch.export.load. PCW does not call these interfaces or torch.load; its supported model readers use safetensors. This limits the affected entry points in PCW, but does not make arbitrary use of an affected PyTorch runtime safe.

See the runtime advisory review for affected profiles, evidence, and the sparse GPU upgrade path. No dependency alert is dismissed by this review. Do not alter saved run metadata or bypass runtime checks to resume a run under another PyTorch version.