CVE-2026-22773

vLLM is an inference and serving engine for large language models (LLMs). In versions from 0.6.4 to before 0.12.0, users can crash the vLLM engine serving multimodal models that use the Idefics3 vision model implementation by sending a specially crafted 1x1 pixel image. This causes a tensor dimension mismatch that results in an unhandled runtime error, leading to complete server termination. This issue has been patched in version 0.12.0.

An attacker identifies a public-facing API or internal endpoint serving a multimodal LLM application built on vLLM — discoverable via job postings, GitHub repos, or API fingerprinting. With zero credentials required, the attacker constructs a multipart HTTP POST request containing a 1x1 pixel PNG image and submits it to any inference endpoint processing images via Idefics3. The vLLM server attempts to process the image, encounters a tensor dimension mismatch on the anomalous image shape, throws an unhandled runtime exception, and terminates the entire server process. Total exploit complexity: generate a 1x1 image (trivially done with any image library or even manually), send one HTTP request. The attacker can repeat this loop to maintain a sustained denial-of-service condition against any unpatched deployment.