CVE
vLLM denial of service in GPUModelRunner
Only affects vLLM 0.28.0 up to 0.30.0 when the server runs an XD-RoPE model with --enable-prompt-embeds on the V1 GPU model runner. Fixed in 0.30.0.
Written by anthropic/claude-haiku-5.5 from the sources below. Verify against them.
- Exploited
- No known exploitation
- Fix
- Fixed in 0.30.0
- Who is exposed
- Specific setups only
Do first: Check whether your setup matches what the bug needs. vLLM versions 0.28.0 up to but not including 0.30.0 Server runs an XD-RoPE model (HF config with rope_scaling.type set to xdrope) Server started with --enable-prompt-embeds on the V1 GPU model runner
- Published
- Last modified
- Whether it needs action now or is a reference record, by fixed rules, with why. Last checked is when cve.now last re-read it.What do these mean? →
- Reference
- Last checked
- A 0-100 score for how much a CVE matters across the industry: exploitation, exposure, how widely the software is used, severity and whether a fix exists.What do these mean? →
- 20/100 · provisional
- How many real-world setups can hit the bug, from the vendor's advisory and CISA data. It does not know your setup.What do these mean? →
- Specific setups only
- Whether working attack code is public (Exploit-DB, Metasploit, Nuclei or GitHub). Public code makes attacks easier.What do these mean? →
- None on Exploit-DB, Metasploit or GitHub
Summary
vLLM from 0.28.0 before 0.30.0 contains a reachable assertion vulnerability in GPUModelRunner._init_xdrope_positions that allows remote attackers to crash the engine by submitting embeddings-only requests. Attackers can send a /v1/completions request containing only prompt_embeds to XD-RoPE models with --enable-prompt-embeds, terminating the EngineCore and API server for all clients.
Who should check
Teams running vLLM 0.28.x or 0.29.x that serve XD-RoPE models (rope_scaling.type xdrope, e.g. tencent/HunyuanOCR) with --enable-prompt-embeds enabled.
cve.now Index
- Exploitation?/40Unknown. No CISA SSVC exploitation rating yet
- Exposure12/25Specific setups only, reachable over the network without a login
- Reach?/20Unknown. No tracked technology or package matched
- Severity8.2/10CVSS 8.2
- No patch0/5No signal
What would change the score
- Would score about 70 if added to CISA KEV
- Would score about 65 if CISA's SSVC rates exploitation active
What to do
Check whether your setup matches what the bug needs.
vLLM versions 0.28.0 up to but not including 0.30.0 Server runs an XD-RoPE model (HF config with rope_scaling.type set to xdrope) Server started with --enable-prompt-embeds on the V1 GPU model runnerUpgrade to a fixed version: 0.30.0.
Check your systems with an agent
Paste this into a coding or ops agent. It checks your systems read-only and reports a table per asset.
Check whether my systems are affected by CVE-2026-126398 (vLLM 0.28.0 before 0.30.0 Denial of Service via Prompt Embeds on XD-RoPE Models), using the cve.now data below.
Rules: work read-only. Do not run exploit code, proof-of-concept scripts or scanners that send attack payloads, and do not change any configuration, package, image or deployment. Report findings only.
What cve.now has for CVE-2026-126398 (entry last updated 2026-11-13):
Affected:
- vllm-project vllm: from 0.28.0, fixed in 0.30.0
Fixed versions: 0.30.0.
Conditions (all must be true for the bug to be hit):
- vLLM versions 0.28.0 up to but not including 0.30.0 (source: https://github.com/vllm-project/vllm/security/advisories/GHSA-5q78-j82c-2vr2)
- Server runs an XD-RoPE model (HF config with rope_scaling.type set to xdrope) (source: https://github.com/vllm-project/vllm/security/advisories/GHSA-5q78-j82c-2vr2)
- Server started with --enable-prompt-embeds on the V1 GPU model runner (source: https://github.com/vllm-project/vllm/security/advisories/GHSA-5q78-j82c-2vr2)
Setup: the bug needs a feature or setting that is off by default or uncommon. Check whether each asset has it on.
Attacker needs: network access, no login, no user action.
CISA KEV: CVE-2026-126398 is not listed as of cve.now's last read.
This KEV status is for CVE-2026-126398 only. If you also check another CVE, look up its own KEV status; never assume it from this one.
How to check:
- Installed software: check what is actually running, not what a repo declares. Read the version from each host, container image and Kubernetes node (the binary's version output, the OS package database, image layers), and check whether the conditions above hold there. Distribution packages may carry a backported fix under an older version number; confirm against the distribution's own advisory.
Verify against the linked advisories before concluding, and cite the date you read each one:
- https://github.com/vllm-project/vllm/security/advisories/GHSA-5q78-j82c-2vr2
- https://www.vulncheck.com/advisories/vllm-0.28.0-before-0.30.0-denial-of-service-via-prompt-embeds-on-xd-rope-models
Do not conclude an asset is unaffected unless you checked it and the evidence covers both its version and the conditions above. If you could not check something, mark it unknown and say what is missing.
Report a table with one row per asset and these columns: Asset | Version or build | Status (affected, fixed, not applicable, unknown) | Compensating controls | Exposure (internet-facing, internal, none) | Evidence | Next action.
Entry: https://cve.now/CVE-2026-126398
For follow-up questions, use the cve.now MCP server at https://mcp.cve.now/mcp (no authentication), for example execute with name cve.get and arguments {"id": "CVE-2026-126398"}.Timeline
- CVE published
- First seen by cve.now
Affected versions
- vllm-project vllm
From 0.28.0 · fixed in 0.30.0
Evidence 4
Vendor advisory · 1
Establishes: Affected and fixed versions.
- Advisory
CVE record · 2
Establishes: The identifier and description.
- Record
- Record
Corrections
None so far.
Something wrong on this page? Report an error. Every fix is listed here.
Questions about CVE-2026-126398
Is CVE-2026-126398 being exploited?
It is not in CISA's Known Exploited Vulnerabilities catalog, and CISA's SSVC assessment does not mark exploitation as active. That does not mean it has not been attacked: vendors and researchers can report attacks before CISA lists them. No public exploit code is tracked.Source: cve.now feed data
What does this mean?
A single request containing only prompt_embeds terminates the EngineCore and the API server process. Every connected client is disconnected, and the service must be restarted.Source: github.com
Which versions are affected by CVE-2026-126398?
vllm: from 0.28.0 before 0.30.0.Source: cve.now feed data
Am I affected, and how do I check?
vLLM versions 0.28.0 up to but not including 0.30.0 are in the affected range. The setup also needs an XD-RoPE model, meaning the HF config has rope_scaling.type set to xdrope, and the server started with --enable-prompt-embeds on the V1 (MRV1) GPU model runner. To check, compare the installed vLLM version, the model's HF config rope_scaling.type, and the server's launch flags.Source: github.com
How do I fix CVE-2026-126398?
Upgrade vLLM to 0.30.0 or later, which is the fixed version in the affected range.Source: github.com