vLLM Inference Engine Vulnerability Allowing Invalid Sparse Tensor Access

CVE-2026-73557

What is CVE-2026-73557?

The vLLM inference and serving engine for large language models has a vulnerability stemming from the unsafe handling of sparse tensors. Specifically, during the processing of prompt embeddings, a race condition can occur due to concurrent submissions through the AsyncMultiModalItemTracker. This issue arises from the use of torch's tensor state management, allowing invalid sparse tensors to potentially be converted to dense formats, bypassing previous security mechanisms intended to guard against such occurrences. Users are advised to update to version 0.26.0, where this issue has been resolved.

References