CVE-2026-22778 – vLLM Remote Code Execution Vulnerability
“A single malicious media request can turn an AI inference server into an attacker-controlled system.”
This patch addresses a critical remote code execution vulnerability (CVE-2026-22778) in vLLM, a widely used engine for serving large language models. The issue affects vLLM versions 0.8.3 through 0.14.0 and is resolved in vLLM 0.14.1. The vulnerability allows attackers to trigger memory disclosure through malformed image input, leaking heap memory addresses that weaken system protections.
The exposed memory information significantly reduces the effectiveness of ASLR protections, allowing attackers to combine the leak with a heap overflow in the JPEG2000 decoder used by OpenCV/FFmpeg. When chained together, the attack can lead to remote code execution and full server compromise on systems running vulnerable vLLM deployments. The flaw is tracked as CVE-2026-22778, with a CVSS v3.1 score of 9.8 (Critical).
The patch in vLLM 0.14.1 sanitizes error responses from image processing and blocks the exploit chain that could expose internal memory data. Systems using multimodal endpoints that process external images or videos face the highest risk if left unpatched.
Key Details
- Affected Product
- Vllm Vllm
- Attack Vector
- Network
- Attack Complexity
- Low
- Privileges Required
- None
- User Interaction
- None
- CWE Classification
- CWE-532