Known CVEpypi/vllm
CVE-2026-115912: vLLM shared-memory multimodal cache handle forgery reaches pickle.loads()
High
With the shared-memory multimodal processor cache (--mm-processor-cache-type shm), vLLM trusted a caller-supplied address and ID as a handle to shared memory, so an authenticated inference caller could point a worker at bytes it chose and have them unpickled, running Python in the worker, over ordinary HTTP routes and without insecure serialization enabled. Fixed in 0.32.0.
Affected and fixed versions
Affected: < 0.32.0. Fixed in 0.32.0.
Sources
See every known CVE, and the ones in the news, in the Learning Center