vLLM published eight security advisories. All of them close in version 0.32.0.
- CVE-2026-115912 scores 7.5. A forged shared memory handle reaches pickle.loads inside the worker.
- The chain uses three public HTTP routes and works with insecure serialization switched off.
- It applies where the multimodal cache runs as shared memory, the –mm-processor-cache-type shm setting.
- CVE-2026-115915 scores 8.1. An EPS file labelled as an image reached the host Ghostscript.
- CVE-2026-115916 kills the engine when a prompt token id equals the vocabulary size.
- The fix ships in pull request 59357, and the advisory range is every version below 0.32.0.
- Most teams run vLLM inside a platform or a vendor build, where the version string does not say which advisories are closed.
- The research was credited to Patch the Planet, a Trail of Bits and OpenAI collaboration.
Grep every serving manifest for the shared memory cache setting before anything else. Read the full analysis.
Get the next one before it is old news
Independent analysis of cloud-native infrastructure, Kubernetes and data center economics. No vendor spin.
