LiveActive security incident?Get immediate response
CVE Record

CVE-2026-22807: vLLM affected by RCE via auto_map dynamic module loading during model initialization

vLLM is an inference and serving engine for large language models (LLMs). Starting in version 0.10.1 and prior to version 0.14.0, vLLM loads Hugging Face `auto_map` dynamic modules during model resolution without gating on `trust_remote_code`, allowing attacker-controlled Python code in a model repo/path to execute at server startup. An attacker who can influence the model repo/path (local directory or remote Hugging Face repo) can achieve arbitrary code execution on the vLLM host during model load. This happens before any request handling and does not require API access. Version 0.14.0 fixes the issue.

HighCVSS 8.8Not KEV-listedUpdated
Discuss exposureBack to CVEs
Glexia's TakeAutomated analysishigh

Security readout for executives and security teams

Plain-English summary

This flaw lets code hidden in a chosen model repository or local model directory run when vLLM starts loading the model. It matters for organizations using vLLM to serve LLMs because compromise can occur before any API request is made. The sources name vLLM 0.14.0 as the fixed version.

Executive priority

Treat this as a high-priority remediation for any production vLLM service using externally sourced or changeable models. The business risk is host compromise of model-serving infrastructure, not just a bad API response. Prioritize internet-facing, shared, and automated model-loading environments first.

Technical view

vLLM versions 0.10.1 through before 0.14.0 load Hugging Face auto_map dynamic modules during model resolution without enforcing trust_remote_code. If an attacker can influence the model repo or path, attacker-controlled Python can execute on the vLLM host during initialization. The issue is classified as CWE-94 with CVSS 8.8.

Likely exposure

Exposure is likely in deployments running vLLM 0.10.1 to 0.13.x where model references come from operators, CI/CD, configuration, remote Hugging Face repos, or local paths that are not strictly trusted. Public API access is not required, but model selection influence is required.

Exploitation context

The provided sources do not show active exploitation, and KEV is false. Exploitation depends on getting a vulnerable server to initialize a malicious or attacker-controlled model source. The attack occurs during model load, before request handling, so normal API authentication may not help.

Researcher notes

The key control failure is dynamic module loading from auto_map during model resolution without trust_remote_code gating. The affected range and fix are clearly identified in the source bundle. Evidence about exploit-in-the-wild activity is absent, so do not claim active exploitation from these sources.

Mitigation direction

  • Upgrade vLLM to version 0.14.0 or a vendor-fixed downstream build.
  • Do not load untrusted Hugging Face repositories or local model directories.
  • Restrict who and what can change model repo, path, and deployment configuration.
  • Review Red Hat advisories for downstream package status and fixed releases.
  • Run vLLM with least privilege and isolate model-serving hosts where feasible.

Validation and detection

  • Inventory vLLM package and container versions across development, staging, and production.
  • Identify deployments running vLLM 0.10.1 through 0.13.x.
  • List every configured model path, Hugging Face repo, and model selection parameter.
  • Check whether users, CI/CD, or operators can influence loaded model references.
  • Confirm upgraded systems report vLLM 0.14.0 or vendor-fixed builds.
Prepared
Confidence
high
Sources
12

Public sources used

Generated from the cited source records. This long-tail analysis has not been individually reviewed by a named human.

Potential ATT&CK relevance

Conservative CVE-to-ATT&CK context

These mappings and lookup hints may be relevant to the vulnerability behavior, CWE, affected product, or exposure path. Glexia-inferred context is not an official MITRE, ATT&CK, CWE, or CVE Program mapping.

ATT&CK lookup starting points

Use these exact CWE pages and searches to review the Glexia ATT&CK library from this CVE's weakness and description context.

cwe · medium confidence lookup

CWE-94: Code execution behavior lookup

Code execution and unsafe deserialization weaknesses often justify reviewing execution behavior and process telemetry. Open the exact CWE lookup page first, then review the ATT&CK searches from that MITRE weakness context. This is a Glexia lookup hint, not an official ATT&CK mapping.

Open ATT&CK lookup
description · low confidence lookup

Execution behavior lookup

The CVE wording references code or command execution, so execution technique review may help defensive triage. This is a Glexia inferred lookup path, not an official MITRE, ATT&CK, or CVE Program mapping.

Open ATT&CK lookup
cve · low confidence lookup

CVE-2026-22807 mapping review

Open the CVE-to-ATT&CK bridge for reviewed, inferred, or future official mappings tied to this CVE.

Open ATT&CK lookup
Vulnerability profileCVE Program record
Severity
High
CVSS
8.8 (3.1)
Known Exploited
No
Published

Vector: CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H

Official CVE source material

CNA and ADP enrichment extracted from CVE v5

These fields come from the CVE record and ADP containers, not from Glexia's Take. They preserve time-varying source decisions such as CISA SSVC, KEV status, CVSS metrics, and provider references.

2CVSS vectors
5Timeline events
2ADP providers
17Source links

SSVC decision data

CISA-ADPCISA Coordinator
Timestamp
Version
2.0.3
Exploitation: noneAutomatable: noTechnical Impact: total

CVSS vector scores

2 official scores

We collect every scored CVSS vector available in the official CNA and ADP containers. When more than one version is present, the table keeps the source vectors side by side instead of collapsing them into the highest score.

ScoreVersionSeverityVectorExploitImpactSource
8.8CVSS 3.1HighCVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H2.85.9GitHub_M
8.8CVSS 3.1HighCVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H2.85.9redhat-SADP

Vulnerability scoring details

Base CVSS 3.1 score

8.8High
CVSS 3.1 vector shape for CVE-2026-22807Attack VectorAttack ComplexityPrivileges RequiredUser InteractionScopeConfidentiality ImpactIntegrity ImpactAvailability Impact

Vector: CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H

Attack Vector
NetworkAdjacentLocalPhysical
Attack Complexity
LowHigh
Privileges Required
NoneLowHigh
User Interaction
NoneRequired
Scope
ChangedUnchanged
Confidentiality Impact
HighLowNone
Integrity Impact
HighLowNone
Availability Impact
HighLowNone

Vulnerability timeline

Timeline events are normalized from CVE metadata, CNA source timelines, ADP timelines, and KEV metadata when present.

  1. CVE reservedCVE Program

    The CVE ID was reserved by the assigning CNA.

  2. ADP timelineredhat-SADP

    Made public.

  3. CVE publishedCVE Program

    The CVE record was published.

  4. ADP timelineredhat-SADP

    Reported to Red Hat.

  5. CVE updatedCVE Program

    The CVE record metadata indicates this as the latest update time.

ADP provider summaries

CISA-ADPCISA ADP Vulnrichment
other:ssvc
redhat-SADPvLLM: vLLM: Arbitrary code execution via untrusted model loading
other:Red Hat severity ratingcvssV3_1
  • 2026-01-21T22:00:55.823Z: Reported to Red Hat.
  • 2026-01-21T21:13:11.894Z: Made public.

Source materials

Affected products

Products and packages named in the record

VendorProductVersion / packageStatus
vllm-projectvllm>= 0.10.1, < 0.14.0Listed
Weakness

CWE details

CWE-94

CWE links open Glexia weakness intelligence pages with official CWE context, developer remediation guidance, and related CVE mappings.

CWE-94 · source CWE mapping

Improper Control of Generation of Code ('Code Injection')

Improper Control of Generation of Code ('Code Injection') represents a recurring weakness pattern that can create exploitable paths when design, validation, or implementation controls are missing.

Open CWE intelligenceSame-CWE CVEs