CVE-2025-33225 / CVE-2025-33235: Predictable Log Files and Race Conditions in NVIDIA’s Training Resilience Layer
| CVE | Component | Type | Disclosed |
|---|---|---|---|
| CVE-2025-33225 | Resiliency Extension — log aggregation | Predictable log-file names | December 12, 2025 |
| CVE-2025-33235 | Resiliency Extension — checkpointing core | Race condition | December 12, 2025 |
Discovered by naxusaireporting.
For executives
NVIDIA Resiliency Extension (NVRx) handles fault tolerance, checkpoint saving, and training recovery in large-scale AI training runs. In December 2025, two vulnerabilities were disclosed. The first involves predictable log file names that allow an attacker to pre-position a file at a known path and cause the logging system to write to attacker-controlled locations. The second is a race condition in the checkpointing core that an attacker can exploit to tamper with training state or escalate privileges. NVRx is used in NVIDIA's NeMo Framework — any organisation running NeMo for large-scale model training has NVRx in their stack.
CVE-2025-33225: predictable log-file names
The naming scheme is predictable — an attacker who understands NVRx's log file naming convention can determine, before the training run begins, what file paths NVRx will attempt to write logs to. With that knowledge, the attacker pre-creates files at those paths, directing log output to security-relevant locations on the filesystem.
CVE-2025-33235: race condition in the checkpointing core
NVRx's checkpointing mechanism forks a separate process to handle checkpoint I/O. CVE-2025-33235 is a race condition in this handoff. Race conditions in checkpoint management are particularly concerning because checkpoints are the ground truth of a training run's progress. A compromised checkpoint that is subsequently loaded produces a model whose training history has been adversarially influenced.
Patch
Upgrade NVIDIA Resiliency Extension to version 0.5.0 or later.