# CVE-2026-94627: vLLM through 0.29.0 GPU KV Cache Leak via Mooncake Transfer ID Collision

> Attackers can exhaust GPU memory in vLLM deployments using Mooncake connectors by triggering KV cache leaks, causing service denial.

- **Severity:** High 8.7 (`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`)
- **Exploited in the wild:** Not known
- **Login needed:** Not needed
- **Attackable:** Over the internet
- **Victim action needed:** Not needed
- **Fix available:** Yes
- **Published:** 21 Sep 2026
- **Updated:** 22 Sep 2026
- **Also known as:** GHSA-pjfg-w894-9457, MINI-3mm4-r5p9-qq86, MINI-w4fc-r7wx-gwvx, MINI-xqjq-8579-f29c

## What is CVE-2026-94627?

CVE-2026-94627 is a high-severity denial of service vulnerability in Vllm Vllm, published on 21 Sep 2026. An attacker can exploit it over the internet without logging in, with no action from the victim.

### Official 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.

Weakness: CWE-401 (Missing Release of Memory after Effective Lifetime)

## Which versions are affected by CVE-2026-94627?

It affects 1 product.


Products: Vllm Vllm.

## How do I fix CVE-2026-94627?

Apply the vendor's patch.

## Is CVE-2026-94627 being exploited?

There is no known exploitation of CVE-2026-94627 in the wild and no public exploit code.

## Other vulnerabilities in Vllm Vllm

- [CVE-2025-32444](https://insights.sbomify.com/CVE-2025-32444/): vLLM Vulnerable to Remote Code Execution via Mooncake Integration (critical 10.0)
- [CVE-2026-22778](https://insights.sbomify.com/CVE-2026-22778/): vLLM leaks a heap address when PIL throws an error (critical 9.8)
- [CVE-2025-47277](https://insights.sbomify.com/CVE-2025-47277/): vLLM Allows Remote Code Execution via PyNcclPipe Communication Service (critical 9.8)
- [CVE-2024-11041](https://insights.sbomify.com/CVE-2024-11041/): Remote Code Execution in vllm-project/vllm (critical 9.8)
- [CVE-2026-48746](https://insights.sbomify.com/CVE-2026-48746/): vLLM: OpenAI auth bypass (critical 9.1)

## Fix prompt for AI coding agents

```text
Check this codebase for CVE-2026-94627 and fix it if it is affected.

What it is: CVE-2026-94627 is a high-severity denial of service vulnerability in Vllm Vllm, published on 21 Sep 2026. An attacker can exploit it over the internet without logging in, with no action from the victim.
Exploitation: There is no known exploitation of CVE-2026-94627 in the wild and no public exploit code.

Affected product: Vllm Vllm. No package-level version data is available, so check the vendor advisory linked from the reference page.

Task:
1. Search the dependency manifests and lockfiles (for example package.json, package-lock.json, requirements.txt, poetry.lock, uv.lock, pom.xml, build.gradle, go.mod, Cargo.lock, Gemfile.lock) and any Dockerfiles for the affected packages, including transitive dependencies.
2. If an affected version is in use, upgrade it to the fixed version for its release line with the project's package manager, so the lockfile is regenerated. Prefer the smallest upgrade that fixes it.
3. If there is no fix, or the upgrade would break the build, do not guess: explain the risk and the safest workaround instead.
4. Run the tests and summarise what changed and why.
5. If nothing in the project is affected, say so and change nothing.

Reference: https://insights.sbomify.com/CVE-2026-94627/ (Markdown: https://insights.sbomify.com/CVE-2026-94627/index.md)
```

## Sources

- CVE record: https://www.cve.org/CVERecord?id=CVE-2026-94627
- NVD: https://nvd.nist.gov/vuln/detail/CVE-2026-94627
- OSV: https://osv.dev/vulnerability/CVE-2026-94627

---
From sbomify Insights: https://insights.sbomify.com/CVE-2026-94627/
Summaries are generated and may contain mistakes; the official description above is the record.
