vLLM Flaw (CVE-2025-62164) Risks Remote Code Execution via Malicious Prompt Embeddings

SecurityOnline[.]info reports a high-severity vulnerability in vLLM, tracked as CVE-2025-62164, which could allow remote code execution via malicious prompt embeddings. The flaw affects vLLM versions 0.10.2 and later and is rated CVSS 8.8, potentially enabling attackers to crash servers or run arbitrary code through the Completions API endpoint.

According to the advisory, a memory corruption vulnerability arises during deserialization of user-supplied embeddings, with the code path loading tensors via torch[.]load(tensor, weights_only=True). Due to a change in PyTorch 2.8.0, sparse tensor integrity checks are disabled by default, meaning maliciously crafted tensors can bypass internal bounds checks and trigger an out-of-bounds memory write during to_dense().

The patch, merged under pull request #27204, prompts organisations running vLLM to upgrade immediately and audit any external-facing model-serving interfaces.