vLLM vulnerabilities

A vulnerability in vLLM (versions 0.10.2 to before 0.11.1) in the Completions API endpoint. When processing user-supplied prompt embeddings, the server loads serialized tensors via torch.load() without sufficient validation. Due to a change in PyTorch 2.8.0, sparse tensor integrity checks are disabled by default, allowing maliciously crafted tensors to bypass bounds checks and trigger an out-of-bounds memory write during to_dense(), leading to a crash (DoS) and potentially remote code execution (RCE).

vLLM before version 0.11.0rc2 validates API keys using a method vulnerable to a timing attack, as string comparison takes longer the more characters are correct. Analysis across many attempts could allow an attacker to determine the next correct character and bypass authentication.