CVE-2026-94622Patch(vllm / vllm)

LOWCVSS 8.7 · HIGH

Signal is active with 1 mentions in latest observed window

Immediate actions

Patch vllm vllm systems immediately

Recommended action window: Monitor and triage in normal cycle

NVD description

vLLM versions through 0.29.0 contain a denial of service vulnerability in the NIXL connector's metadata handling for prefill/decode disaggregated deployments. Attackers can send requests with incomplete kv_transfer_params dictionary entries to trigger an uncaught KeyError in EngineCore scheduling, causing the decode engine to terminate and making all routed requests fail until manual restart.

vllm CVEs Back to dashboard

0.5/ 10 priority

Sources & remediation

Patch / fix

https://github.com/vllm-project/vllm/pull/54807

Vendor / third-party advisories

https://www.vulncheck.com/advisories/vllm-through-0.29.0-denial-of-service-via-incomplete-nixl-kv-transfer-metadata

Other references

Weakness type (CWE)

CWE-248

Priority

LOW

Exploitation

NONE

PoC

NONE

Patch

AVAILABLE

Momentum

STABLE

Are you affected?

If you run products in this scope, you should treat this CVE as relevant to your environment.

vllm

Threat summary

Patch or workaround signal is available
3 mentions across 3 observed days
Momentum state: stable

What's happening

Patch or workaround mentioned in 2 signals
Technical details provided in 3 signals
Disclosure: 1 classified signal
Peaked 2d ago at 1 mentions (2026-09-22); latest day: 1
3 total mentions across 3 days

Affected systems

Vendors

vllm

Products

vllm

Deep dive

Activity timeline3 mentions / 3d

Signal classification2 categories

Patch

266.7%

Disclosure

133.3%

Referenced assets3 URLs

By indicator

Classification over time

Date	Total	Labels

2026-09-22	1	Disclosure1
2026-09-24	1	Patch1
2026-10-05	1	Patch1

Full discourse3 posts

Tesla Fan Dan@Tesla_FANtastic44d
Patch
vLLM shipped version 0.30.0 on September 22, 2026. If you use vLLM to serve models, that is the version to pin right now. On September 21, CVE-2026-94622 was published for versions through 0.29.0. In plain terms: on setups that split work with NIXL (prefill on one side, decode on another), bad or incomplete metadata can knock over the decode side until you restart it. That is a availability problem, not a reason to invent extra version numbers. The verified upgrade target is the GitHub release v0.30.0 (Docker image vllm/vllm-openai:v0.30.0). Do not wait on an unverified mid-patch name. If you do not run vLLM, this post is not for you. Do today: Check which vLLM version you are running. If it is 0.29.0 or older and you serve real traffic, upgrade to 0.30.0 and confirm the version before you send traffic back. https://github.com/vllm-project/vllm/releases/tag/v0.30.0
Post summary
vLLM 0.30.0 patches CVE-2026-94622 (versions ≤0.29.0); a metadata issue in NIXL-split prefill/decode setups can crash the decode side. The text directs users to upgrade to v0.30.0 as the verified fix.
1001063
13 followersView on X
Hugo | DevOps | Cybersecurity 🇱🇻@HugoValters33d
Patch
CVE-2026-94622: DoS in vLLM (through 0.29.0). Crafted kv_transfer_params crash the NIXL connector, killing the decode engine until manual restart. CVSS 7.5. No patch yet - track it and restart capacity now. https://www.valtersit.com/cve/CVE-2026-94622 #CVE #infosec #vLLM #vLLM #CVEALERT #CVE #infosec #Linux #XSS #valtersit #snippet #SysAdmin #cybersecurity #devsecops #devops #developer #100daysofcode #git #github #gitlab #redteam #blueteam #ethicalhacker #cybersecurityawareness #cybersecuritynews #cybersecuritytips #python #hacker #kali #ubuntu #debian #docker
Post summary
The tweet discloses a DoS vulnerability (CVE-2026-94622) in vLLM versions up to 0.29.0 with technical details and a CVSS 7.5 score, explicitly noting no patch exists yet but recommending a manual restart workaround to restore the decode engine.
0000028
1.0K followersView on X
ThreatAft@ThreatAft46d
Disclosure
🚨 vLLM DoS CLUSTER — 5 × CVSS 7.5 One crafted request crashes the shared inference worker. Every client loses service. • CVE-2026-94622 — NIXL metadata • CVE-2026-94623 — Prefix caching → https://threataft.com/articles/vllm-dos-cluster-five-memory-exhaustion-cves #vLLM #CVE #LLM #PatchNow #CyberSecurity #ThreatIntel
Post summary
The tweet announces a cluster of vLLM denial-of-service vulnerabilities, providing CVSS scores and affected components, but does not include PoC, exploit code, or patch details.
0000062
43 followersView on X

CPE platform detail1 entries

Part:

Vendor:

Product:

Version:

Target SW:

Target HW:

1 of 1 entries

Part	Vendor	Product	Version	Target SW	Target HW
App	vllm	vllm	-	-	-

Explore more

Back to dashboard vllm CVEs vllm vulnerabilities

What researchers are saying

CVE-2026-94622

Tesla Fan Dan[verified]@Tesla_FANtasticPatch

vLLM 0.30.0 patches CVE-2026-94622 (versions ≤0.29.0); a metadata issue in NIXL-split prefill/decode setups can crash the decode side. The text directs users to upgrade to v0.30.0 as the verified fix.

Hugo | DevOps | Cybersecurity 🇱🇻[verified]@HugoValtersPatch

The tweet discloses a DoS vulnerability (CVE-2026-94622) in vLLM versions up to 0.29.0 with technical details and a CVSS 7.5 score, explicitly noting no patch exists yet but recommending a manual restart workaround to restore the decode engine.

ThreatAft@ThreatAftDisclosure

The tweet announces a cluster of vLLM denial-of-service vulnerabilities, providing CVSS scores and affected components, but does not include PoC, exploit code, or patch details.