CVE-2026-94626

vLLM through 0.29.0 Memory Exhaustion via Unvalidated NIXL tp_size

Published Sep 21, 2026 CVSS 4.0 · 8.7 High EPSS < 1% KEV · Not listed
Description
vLLM through 0.29.0 fails to validate the tp_size parameter in kv_transfer_params on OpenAI-compatible completion endpoints, allowing attackers to allocate unbounded memory. Attackers can supply arbitrary tp_size values in prefill/decode disaggregated deployments to exhaust memory and trigger kernel OOM-kill of the decode worker process.

Analysis

AI-generated

Technical impact

vLLM versions up to and including 0.29.0 contain a flaw that fails to validate the tp_size parameter on OpenAI-compatible completion endpoints. An attacker can supply arbitrarily large tp_size values in kv_transfer_params, causing the decoder worker to allocate that amount of memory and exhaust the host system’s RAM. The uncontrolled memory allocation aligns with CWE‑770 and CWE‑789 weaknesses and can trigger the kernel OOM‑killer to terminate the offending process, resulting in a denial‑of‑service for the affected deployment.

Affected Systems

The vulnerability affects the vllm-project vllm product. All releases through 0.29.0 are impacted. The flaw is triggered by the OpenAI‑compatible completion endpoints exposed by the service.

Risk and Exploitability

The CVSS score of 8.7 categorizes this as high severity. The EPSS score of < 1 % indicates that active exploitation is unlikely at this time. The vulnerability is not listed in CISA KEV. Attackers can remotely exploit the exposed completion API by supplying arbitrary tp_size values, which the server accepts without bounds. When such a request reaches the server, the memory allocation is almost deterministic; the OOM killer will terminate the decoder worker, potentially disrupting all concurrent decoding activities until the process is restarted or the system is scaled out.

Remediation

No solution or workaround provided in the CVE record.

OpenCVE Recommended Actions

  • Update vllm to the latest release where tp_size validation has been added
  • Configure the API gateway or load balancer to reject or cap tp_size values to a safe maximum and apply rate limiting on completion requests
  • Set container or VM memory limits, enable OOM protection, and monitor worker process memory usage to restart or scale out when high usage is detected

Generated by OpenCVE AI on September 23, 2026 at 02:06 UTC.

Tracking

Sign in to view the affected projects.

Advisories

No advisories available
History
Type Values Removed Values Added
Metrics ssvc

{'options': {'Automatable': 'yes', 'Exploitation': 'none', 'Technical Impact': 'partial'}, 'version': '2.0.3'}

Type Values Removed Values Added
Description vLLM through 0.29.0 fails to validate the tp_size parameter in kv_transfer_params on OpenAI-compatible completion endpoints, allowing attackers to allocate unbounded memory. Attackers can supply arbitrary tp_size values in prefill/decode disaggregated deployments to exhaust memory and trigger kernel OOM-kill of the decode worker process.
Title vLLM through 0.29.0 Memory Exhaustion via Unvalidated NIXL tp_size
First Time appeared Vllm
Vllm vllm
Weaknesses CWE-789
CPEs cpe:2.3:a:vllm:vllm:*:*:*:*:*:*:*:*
Vendors & Products Vllm
Vllm vllm
References
Metrics cvssV3_1

{'score': 7.5, 'vector': 'CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H'}

cvssV4_0

{'score': 8.7, 'vector': 'CVSS:4.0/AV:N/AC:L/AT:N/PR:N/UI:N/VC:N/VI:N/VA:H/SC:N/SI:N/SA:N'}

Weaknesses
  • CWE-770

    Allocation of Resources Without Limits or Throttling

  • CWE-789

    Memory Allocation with Excessive Size Value

Sources

Status
PUBLISHED
Assigner
VulnCheck
Published
Updated
Sep 24, 2026, 10:54 PM
Reserved
Sep 21, 2026, 9:42 PM
Status
Analyzed
Published
Sep 21, 2026, 10:17 PM
Modified
Sep 29, 2026, 12:46 PM
Severity
Important
Public date
Sep 21, 2026, 10:04 PM