chore: document CVE-2026-31221 pytorch-lightning insecure deserialization (no patch available)

[Copilot]

pytorch-lightning==2.0.3 is flagged by Dependabot for CVE-2026-31221 / GHSA-75m9-98v2-hjpm: LightningModule.load_from_checkpoint() calls torch.load() without weights_only=True, enabling arbitrary pickle deserialization from a malicious checkpoint file. No patched version exists yet.

Reachability

Not reachable — high confidence. load_from_checkpoint is never called in this codebase. All checkpoint loading is done manually:

# trainer/__init__.py, inference/infer_meshgpt.py, train_transformer.py
state_dict = torch.load(resume, map_location="cpu")["state_dict"]
model.load_state_dict(get_parameters_from_state_dict(state_dict, "model"))

trainer.fit(model, ckpt_path=...) uses Lightning's internal _load_checkpoint path, which is distinct from the vulnerable load_from_checkpoint API.

Changes

• requirements.txt — added an inline comment on the pytorch-lightning pin referencing the CVE, explaining that no patched version is available and flagging it for update once a fix is released.

Version bump was not performed: the advisory's patched_version field is empty, and the dependency graph is tightly pinned — an untargeted major-version upgrade carries meaningful breakage risk with no clear security benefit given the API is unreachable.

Original prompt

---

This section details the Dependabot vulnerability alert you should resolve

<alert_title>PyTorch Lightning load_from_checkpoint has an insecure checkpoint deserialization</alert_title>
<alert_description>PyTorch-Lightning versions 2.6.0 and earlier contain an insecure deserialization vulnerability (CWE-502) in the checkpoint loading mechanism. The LightningModule.load_from_checkpoint() method, which is commonly used to load saved model states, internally calls torch.load() without setting the security-restrictive weights_only=True parameter. This default behavior allows the deserialization of arbitrary Python objects via the Pickle module. A remote attacker can exploit this by providing a maliciously crafted checkpoint file, leading to arbitrary code execution on the victim's system when the file is loaded.</alert_description>

high
GHSA-75m9-98v2-hjpm, CVE-2026-31221
pytorch-lightning
pip
<vulnerable_versions>= 2.0.3</vulnerable_versions>
<patched_version></patched_version>
<manifest_path>requirements.txt</manifest_path>

https://nvd.nist.gov/vuln/detail/CVE-2026-31221 https://github.com/Lightning-AI/pytorch-lightning https://www.notion.so/CVE-2026-31221-35d1e1393188815f8db7c4fd08076639 https://github.com/advisories/GHSA-75m9-98v2-hjpm

<task_instructions>Resolve this alert by updating the affected package to a non-vulnerable version. Prefer the lowest non-vulnerable version (see the patched_version field above) over the latest to minimize breaking changes. Include a Reachability Assessment section in the PR description. Review the alert_description field to understand which APIs, features, or configurations are affected, then search the codebase for usage of those specific items. If the vulnerable code path is reachable, explain how (which files, APIs, or call sites use the affected functionality) and note that the codebase is actively exposed to this vulnerability. If the vulnerable code path is not reachable, explain why (e.g. the affected API is never called, the vulnerable configuration is not used) and note that the update is primarily to satisfy vulnerability scanners rather than to address an active risk. If the advisory is too vague to determine reachability (e.g. 'improper input validation' with no specific API named), state that reachability could not be determined and explain why. Include a confidence level in the reachability assessment (e.g. high confidence if the advisory names a specific API and you confirmed it is or is not called, low confidence if the usage is indirect and hard to trace). If no patched version is available, check the alert_description field for a Workarounds section — the advisory may describe configuration changes or usage patterns that mitigate the vulnerability without a version update. If a workaround is available, apply it and leave a code comment referencing the advisory identifier explaining it is a temporary mitigation. If neither a patch nor a workaround is available, explain in the PR description why the alert cannot be resolved automatically so a human reviewer can take over. Inspect the repository to determine which package manager is used (e.g. lock files, config files, build scripts) and use that tooling to perform the update — do not edit lock files directly. If the version constraint in the manifest (e.g. package.json, Gemfile, pyproject.toml) caps the version below the fix, update the constraint first. For transitive dependencies, determine whether it is simpler to update the direct dependency that pulls in the vulnerable package or to update the transitive dependency directly, and choose the least disruptive approach. If upgrading to fix the vulnerability forces a major version bump or known breaking changes, review the changelog or release notes, then audit the codebase for usage of affected APIs and fix any breaking changes that are found. If the package manager fails to resolve dependencies (e.g. peer dependency conflicts, incompatible engine constraints), document the error in the PR description rather than attempting increasingly complex workarounds. After updating, check the lock file to confirm the package no longer resolves to a version in the vulnerable range. Keep changes minimal and tightly scoped. Ensure tests, build, type checking, and linting all pass after your changes. If there are any test, lint, or typechecking failures, investigate whether they are caused by the update and fix them if so — do not leave broken tests in the PR. If they were already present before the update, note them in the PR description so a human reviewer can assess whether they are related.</task_instructions>

• Resolves audi/MeshGPT alert #28