Security context

Highlights

Command Injection: 2 prior fixes. Scrutinize any change in this area.

tools/uvr5/webui.py: most-fixed (2 issues). Treat as high-risk during review.

4 high-severity fixes in this history; regressions here are high-impact.

Recurring patterns

Command Injection: Unsanitized input containing shell metacharacters can propagate directly to the shell via os.system when invoking external CLI commands like ffmpeg. Developers should avoid using shell invocation APIs entirely, preferring native Python audio processing libraries (e.g., soundfile and librosa) to fully eliminate the injection vector.

Deserialization of Untrusted Data: Using torch.load to deserialize arbitrary model weights allows arbitrary code execution via Python's underlying pickle implementation. To secure model ingestion, any call to torch.load must enforce weights_only=True or use safe alternative formats to prevent unaligned payload execution.

Denial of Service: Concurrent access to BERT feature extraction can cause race conditions and process crashes when multi-threaded requests are dispatched without synchronization locks. Enforcing thread safety using mutex locks around shared model access blocks is required.