PatchTheCVECVE-2026-94627: vLLM Mooncake connector through 0.29.0 fails to properly manage GPU KV cache block ownership when concurrent c
CVE-2026-94627 is a high-severity flaw (CVSS 7.5) affecting vllm. Below: why it ranks where it does, which versions are affected, what to do, and the exact commands.
vLLM Mooncake connector through 0.29.0 fails to properly manage GPU KV cache block ownership when concurrent child requests share a single transfer ID in prefill/decode disaggregated deployments. Attackers can trigger GPU memory exhaustion by submitting completion requests with multiple prompts, causing orphaned KV cache blocks to accumulate until process restart and eventually preventing legitimate requests from executing.
Why this priority
- High severity
- Remotely exploitable, no login, no user action
A fixed release or distro patch exists.
Risk facts
| CVSS score | 7.5 (High) |
| Attack vector | Network (remote) |
| Attack complexity | Low |
| Privileges needed | None |
| User interaction | None |
| EPSS exploit chance | 0.6% (top 50%) |
| CISA exploited list | No |
| Weakness | CWE-401 |
| Published | 2026-09-21 |
| Last modified | 2026-09-29 |
Last updated 2026-09-29 · Sources: CVE Program, NVD, CISA KEV, EPSS and vendor advisories · How we score
Am I affected?
Timeline
- Published CVE record published
- Updated Record last modified
Scores and fixes here are computed automatically from those and other public sources. Confirm with the vendor advisory before changing production systems.
Quick answers
How do I fix CVE-2026-94627?
vllm: upgrade to newer than 0.29.0. Exact commands for each OS are on this page.
Is CVE-2026-94627 being exploited?
CVE-2026-94627 is not on the CISA Known Exploited Vulnerabilities list at the time of the last update. Check again, as this can change.
What does CVE-2026-94627 affect?
CVE-2026-94627 affects vllm. Severity is high (CVSS 7.5). vLLM Mooncake connector through 0.29.0 fails to properly manage GPU KV cache block ownership when concurrent child requests share a single transfer ID in prefill/decode disaggregated deployments. Attackers can trigger GPU memory exhaustion
How urgent is CVE-2026-94627?
Priority 44/100. High severity; Remotely exploitable, no login, no user action
Related CVEs
- CVE-2026-5497 HIGH 7.5 UPGRADE vLLM versions 0.8.0 and later are vulnerable to an Out-of-Memory (OOM) Denial of Service (DoS) attack due to u
- CVE-2026-48746 CRITICAL 9.1 UPGRADE vLLM is an inference and serving engine for large language models (LLMs)
- CVE-2026-54232 HIGH 8.8 UPGRADE vLLM is an inference and serving engine for large language models (LLMs)
- CVE-2026-41523 HIGH 7.5 UPGRADE vLLM is an inference and serving engine for large language models (LLMs)
- CVE-2026-54236 MEDIUM 5.3 UPGRADE vLLM is an inference and serving engine for large language models (LLMs)