Security context

What an agent needs to avoid regressing past fixes and find the next vuln in this repo.

nvidia-nemo/nemo
main @ dcd7153
14
Fixes
33
CVEs
CRITICAL
Peak severity
41.7%
Coverage
Highlights
Path Traversal: 6 prior fixes. Scrutinize any change in this area.
nemo/core/classes/common.py: most-fixed (3 issues). Treat as high-risk during review.
12 high-severity fixes in this history; regressions here are high-impact.
Recurring patterns

The bug types that recur here, drawn from past fixes, not open vulnerabilities.

Unsafe Deserialization: The use of pickle and wrapping of PyTorch loads (torch.load) allow malicious model weights or datasets to execute arbitrary code upon deserialization. Secure alternatives and strict weights_only validation are required.
Code Execution: Dynamically instantiating target classes using Hydra without validation allows remote execution of arbitrary code via malicious configuration payloads. This requires strict allowlist checking for resolved target packages.
Path Traversal: Tar extraction operations (using tarfile.extractall or tarfile.open) consistently failed to validate that target paths remain inside the target directory, exposing the framework to arbitrary file write issues (Tar Slip).