Skip to main content
Menu
CVEs

CVE

vLLM denial of service in multi-image request handling

Affects vLLM 0.6.0 before 0.33.0 serving multimodal models that accept multiple images per request, via /v1/messages, non-Harmony /v1/responses, and equivalent chat-rendering paths. Fixed in 0.33.0.

CVE-2026-130686EUVD-2026-128760GHSA-RQCJ-6434-QFCHPublished Updated Summary written by a model

8.7HighCNA
Exploited
No known exploitation
Fix
Fixed in 0.33.0
Who is exposed
Default setups

Do first: Check every install: default setups are affected.

Summary

vLLM from 0.6.0 before 0.33.0 lacks a request-wide media budget, allowing remote inference clients to exhaust frontend resources by submitting many remote image URLs in one request. Attackers can send a compact request to /v1/messages or /v1/responses that forces downloading and decoding of every image before prompt-length admission, exhausting memory, CPU, and connections.

Who should check

Teams running vLLM >= 0.6.0 and < 0.33.0 that serve multimodal models accepting multiple images via /v1/messages or /v1/responses, including anyone shipping vLLM from PyPI

“Affected versions >= 0.6.0, < 0.33.0”

github.com

cve.now Index

46Pcve.now Index 46 of 100 · methodology v3 · provisional
  • Exploitation0/40No signal
  • Exposure25/25Default setups, CISA rates it automatable
  • Reach12/20vllm: 578.2K downloads a week
  • Severity8.7/10CVSS 8.7
  • No patch0/5No signal

What would change the score

  • Would score about 86 if added to CISA KEV
  • Would score about 81 if CISA's SSVC rates exploitation active

Impact

  • Widely used

    vllm has 578.2k weekly downloads.Source: pypistats.org

    93 packages depend on vllm.Source: deps.dev

What to do

  • Check every install: default setups are affected.

  • Upgrade to a fixed version: 0.33.0.

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-2026-130686 (vLLM 0.6.0 before 0.33.0 Resource Exhaustion via Multi-Image Requests), 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-2026-130686 (entry last updated 2026-11-18):
Affected:
- PyPI package vllm: from 0.6.0, fixed in 0.33.0
Fixed versions: 0.33.0.
Conditions (all must be true for the bug to be hit):
- Model permits multiple images per request (multimodal model, e.g. LLaVA-style processing with no lower model-specific limit) (source: https://github.com/vllm-project/vllm/security/advisories/GHSA-5r3c-5x4v-8gjh)
- Affected entry points include /v1/messages, non-Harmony /v1/responses, and equivalent chat-rendering paths (source: https://github.com/vllm-project/vllm/security/advisories/GHSA-5r3c-5x4v-8gjh)
Setup: the bug is reachable in a default setup; no special feature or setting is needed.
Attacker needs: network access, no login, no user action.
CISA KEV: CVE-2026-130686 is not listed as of cve.now's last read.
This KEV status is for CVE-2026-130686 only. If you also check another CVE, look up its own KEV status; never assume it from this one.

How to check:
- Packages (PyPI): find every lockfile in each project and run the stack.check operation of the cve.now MCP server (execute) on each one. Compare the resolved versions, not the ranges a manifest declares, with the affected versions above. Include transitive dependencies.

Verify against the linked advisories before concluding, and cite the date you read each one:
- https://github.com/vllm-project/vllm/security/advisories/GHSA-5r3c-5x4v-8gjh
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-2026-130686
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-2026-130686"}.

Timeline

  1. CVE published
  2. First seen by cve.now
  3. CISA SSVC assessment

    Exploitation: none · Automatable: yes · Technical impact: partial

  4. Record last updated

Affected versions

  • vllmPyPI

    From 0.6.0 · fixed in 0.33.0

    Fixed in 0.33.0 (released Nov 2, 2026, 16 days before disclosure).

    The latest release, 0.33.0, is not affected.

Evidence 3

KindSource
Vendor advisory 1Establishes: Affected and fixed versions.
AdvisoryGitHub · vllm-project/vllm · Advisory GHSA-5r3c-5x4v-8gjh
CVE record 2Establishes: The identifier and description.
RecordCVE List · CVE-2026-130686
RecordNVD · CVE-2026-130686

Corrections

None so far.

Something wrong on this page? Report an error. Every fix is listed here.

Questions about CVE-2026-130686

Is CVE-2026-130686 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 remote client can send one small request that makes the API frontend use disproportionate memory, CPU, outbound network connections and connection-pool capacity. This can terminate the API process and degrade availability for other clients.Source: github.com

Which versions are affected by CVE-2026-130686?

vllm (PyPI): from 0.6.0 before 0.33.0.Source: cve.now feed data

Am I affected, and how do I check?

vLLM versions from 0.6.0 before 0.33.0 are in range. The setup involves a multimodal model that permits multiple images in one request, reached through /v1/messages, non-Harmony /v1/responses, or equivalent chat-rendering paths. To check, compare the installed vLLM version against that range and review which model is served and whether it accepts multiple images per request.Source: github.com

How do I fix CVE-2026-130686?

Upgrade vLLM to 0.33.0 or later, which is the fixed version listed for this vulnerability.Source: github.com

More CVEs