vLLM (>= 0.10.2, < 0.13.0) lacks sparse tensor validation in multimodal embeddings when prompt-embeds is enabled. Because PyTorch sparse invariant checks are off by default, crafted embedding requests with negative/out-of-bounds indices can crash or exhaust resources (DoS) and may enable out-of-bounds write-what-where memory corruption. Upgrade to 0.13.0.