GHSA-mcmc-2m55-j8jj

### Summary The fix [here](https://github.com/vllm-project/vllm/pull/27204) for CVE-2025-62164 is not sufficient. The fix only disables prompt embeds by default rather than addressing the root cause, so the DoS vulnerability remains when the feature is enabled. ### Details vLLM's pending change attempts to fix the root cause, which is the missing sparse tensor validation. PyTorch (~v2.0) disables sparse tensor validation (specifically, sparse tensor invariants checks) by default for performance reasons. vLLM is adding the sparse tensor validation to ensure indices are valid, non-negative, and within bounds. These checks help catch malformed tensors. ### PoC NA ### Impact Current fix only added a flag to disable/enable prompt embeds, so by default, prompt embeds feature is disabled in vLLM, which stops DoS attacks through the embeddings. However, It doesn’t address the problem when the flag is enabled and there is still potential for DoS attacks. ### Changes * https://github.com/vllm-project/vllm/pull/30649

An attacker with API credentials to a vLLM-backed inference endpoint — a compromised service account, a malicious insider, or a tenant in a multi-tenant deployment — constructs a request containing prompt embeddings with malformed PyTorch sparse tensors: indices that are negative, out-of-bounds, or violate sparse tensor invariants. Since PyTorch v2.0 disabled these invariant checks by default for performance, vLLM (without the real fix) passes the tensor to the compute path unchecked. This triggers an out-of-bounds memory write, causing the inference worker process to crash (DoS) or potentially corrupt adjacent memory. Because vLLM is typically deployed as a shared inference backend serving multiple users or services, a single malformed request takes down availability for all consumers.