PyTorch Lightning through 2.6.5, fixed in commit d710d68, contains a remote code execution vulnerability in the _load_state function that imports and executes attacker-controlled module names from checkpoint _instantiator hyperparameters. Attackers can craft malicious checkpoint files that bypass weights_only=True protections to execute arbitrary code when LightningModule.load_from_checkpoint is called.
References
| Link | Resource |
|---|---|
| https://github.com/Lightning-AI/pytorch-lightning/commit/d710d689510d50e800f53b3cd773cbca20b1f86f | Patch |
| https://github.com/Lightning-AI/pytorch-lightning/issues/21822 | Exploit Issue Tracking Mitigation |
| https://github.com/Lightning-AI/pytorch-lightning/pull/21832 | Issue Tracking Patch |
| https://www.vulncheck.com/advisories/pytorch-lightning-arbitrary-code-execution-via-instantiator-hyperparameter | Exploit Patch Third Party Advisory |
Configurations
History
No history.
Information
Published : 2026-07-15 18:16
Updated : 2026-08-06 17:31
NVD link : CVE-2026-58659
Mitre link : CVE-2026-58659
CVE.ORG link : CVE-2026-58659
JSON object : View
Products Affected
lightningai
- pytorch_lightning
CWE
CWE-470
Use of Externally-Controlled Input to Select Classes or Code ('Unsafe Reflection')
