CVE-2026-93592Disclosure(vllm / vllm)

🚨HIGH - vLLM DoS via Negative Token IDs in Embeddings/Pooling (CVE-2026-93592) vLLM’s OpenAI-compatible API /v1/embeddings and /pooling fail to validate the lower bound of token IDs; unauthenticated attackers can submit negative IDs to hit a CUDA device-side assert, crashing/poisoning the GPU context. Subsequent requests fail until the process restarts. 👉Affected: vllm < 0.28.0 | Upgrade to 0.28.0

NewNormal Security turns CVEs since the previous batch into new detections, every day. 𝗗𝗮𝗶𝗹𝘆 𝗖𝗩𝗘 𝗥𝗲𝗽𝗼𝗿𝘁 — 19 Sep 2026 𝗔𝗱𝗱𝗲𝗱 to NewScan 𝘁𝗼𝗱𝗮𝘆: 🖥️ Container registry open to anonymous push, or running a build whose bearer-token check can be bypassed — every image, and a tag they can overwrite (zot CVE-2026-61833) 🖥️ Database console on the internet — a login in front of a database whose permission checks any account can walk around (ArcadeDB CVE-2026-93593, CVE-2026-93594, CVE-2026-93595, CVE-2026-93598) 🖥️ Self-hosted AI model server with no authentication — free inference on your GPUs, and one bad request stops the engine for everyone (as seen in vLLM CVE-2026-93592) Test your stack with NewScan — free, self-hosted: https://newnormalsecurity.com/newscan?utm_source=x&utm_medium=social&utm_campaign=daily-cve #infosec #AppSec #ExposedInterface #CSO #REDTEAM