vllm-project · vLLM · published 2026-05-12
CVE-2026-44223vLLM: extract_hidden_states speculative decoding crashes server on any request with penalty parameters
This month Public proof-of-concept exists (SSVC exploitation: poc).
CVE-2026-44223 has a public proof-of-concept exploit, but no confirmed exploitation in the wild. EPSS estimates a 0.4% chance of exploitation in the next 30 days, higher than 31% of all CVEs. This vLLM vulnerability: CVSS 6.5 (v3.1). Fixed upstream in vLLM 0.20.0. Patch in your normal monthly cycle.
Jump to: Remediation Exploit intelAffected versionsDescription
Remediation#
Fixed upstream in vLLM 0.20.0.
Patch events
AI toolingvLLM
vLLM security release: 0.20.0
Upgrade to
0.20.0Why: Public proof-of-concept exists (SSVC exploitation: poc).
Exploit intelligence#
| Signal | Value | Source |
|---|---|---|
| Known exploited | No | CISA KEV |
| Ransomware use | — | CISA KEV |
| EPSS (30-day exploitation probability) | 0.4% (31th pct) | FIRST EPSS v2026.06.15 |
| CVSS base score | 6.5 (v3.1)CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H | CNA / CISA ADP |
| SSVC exploitation | poc | CISA Vulnrichment |
| SSVC automatable | No | CISA Vulnrichment |
| SSVC technical impact | partial | CISA Vulnrichment |
Affected versions (from the CVE record)#
- vllm-project vllm
>= 0.18.0, < 0.20.0(affected)
Description (from the CVE record)#
vLLM is an inference and serving engine for large language models (LLMs). From 0.18.0 to before 0.20.0, the extract_hidden_states speculative decoding proposer in vLLM returns a tensor with an incorrect shape after the first decode step, causing a RuntimeError that crashes the EngineCore process. The crash is triggered when any request in the batch uses sampling penalty parameters (repetition_penalty, frequency_penalty, or presence_penalty). A single request with a penalty parameter (e.g., "repetition_penalty": 1.1) is sufficient to crash the server. This vulnerability is fixed in 0.20.0.
Quoted from the CVE Program record.