High severity · published Sep 21, 2026
CVE-2026-94627
vLLM through 0.29.0 GPU KV Cache Leak via Mooncake Transfer ID Collision
7.5
CVSS
0.6%
EPSS
No
On KEV
Description
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.
CVSS vector
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H
Attack vector
Network
Complexity
Low
Privileges
None
User interaction
None
Confidentiality
None
Integrity
None
Availability
High
Scope
Unchanged
Exploitation
Is it being used?
Exploit prediction (EPSS)0.6% · higher than 50% of CVEs
Exploited in the wildNo evidence yet
CISA KEVNot listed
Ransomware campaign useNot known
Affected products
Which versions
VL
vllm · vllm-project
through 0.29.0 · source: vendor, NVD
References
NVD
https://nvd.nist.gov/vuln/detail/CVE-2026-94627
CVE.org
https://www.cve.org/CVERecord?id=CVE-2026-94627
Product
https://github.com/vllm-project/vllm
Product
https://github.com/vllm-project/vllm/blob/v0.29.0/vllm/distributed/kv_transfer/kv_connector/v1/mooncake/mooncake_connector.py#L1978-L1989
Issue Tracking
https://github.com/vllm-project/vllm/pull/49796
Third Party Advisory
https://www.vulncheck.com/advisories/vllm-through-0.29.0-gpu-kv-cache-leak-via-mooncake-transfer-id-collision
More from vllm
9.8
CVE-2026-115915
vLLM: Untrusted EPS inputs can trigger denial of service or Ghostscript RCE
9.8
CVE-2026-25960
SSRF Protection Bypass in vLLM
9.8
CVE-2026-22778
vLLM leaks a heap address when PIL throws an error
9.8
CVE-2026-22807
vLLM affected by RCE via auto_map dynamic module loading during model initialization
9.8
CVE-2024-9053
vllm-project vllm version 0.6.0 contains a vulnerability in the AsyncEngineRPCServer() RPC server entrypoints