vLLM — memory corruption in Triton bincount kernel via out-of-bounds multimodal audio token indexing

An unbounded array index in vLLM's Triton kernel for multimodal audio processing allows out-of-bounds memory access via crafted token values.

Low individual severity but represents a precise catalogued flaw in vLLM's multimodal audio-handling path worth tracking given the engine's broad deployment.

Prompt token IDs index the penalty prompt-presence bitset in the Triton _bincount_kernel without bounds checking against vocabulary size; attackers can submit multimodal audio requests with tokens equal to vocabulary size to trigger memory corruption.

vLLM through 0.29.0

Upgrade vLLM to the version addressing this issue.