CVE-2026-24779: vllm: SSRF allows internal network access

GHSA-qh4c-xf7m-gxfcHIGHPoC AVAILABLECISA: TRACK*
Published January 27, 2026
CISO Take

vLLM servers processing multimodal inputs are vulnerable to SSRF via a URL parser inconsistency that bypasses host restrictions — any user with inference API access can pivot into your internal network. In Kubernetes-based AI serving stacks (llm-d, KServe), this enables pod-to-pod reconnaissance and service disruption. Patch to vLLM 0.14.1 immediately; if you cannot patch, disable multimodal URL loading at the network perimeter.

What is the risk?

Effective risk is higher than the CVSS 7.1 suggests in AI-native cloud environments. Attack complexity is low once an attacker understands the backslash parsing differential between Python URL libraries — this is documented in the advisory and reproducible. The low-privileges requirement means any user with access to the vLLM inference endpoint (including trial/free-tier customers in multi-tenant deployments) is a potential attacker. In containerized environments with flat internal networking, the blast radius extends well beyond the vLLM pod itself. EPSS is currently near-zero, but the advisory details are public and the technique is straightforward.

What systems are affected?

PackageEcosystemVulnerable RangePatched
vLLMpip—No patch
96.9K93 dependentsPushed 3d ago26% patched~37d to patchFull package profile →
vLLMpip< 0.14.10.14.1
96.9K93 dependentsPushed 3d ago26% patched~37d to patchFull package profile →

How severe is it?

CVSS 3.1
7.1 / 10
EPSS
0.6%
chance of exploitation in 30 days
Higher than 48% of all CVEs
Exploitation Status
Exploit Available
Exploitation: MEDIUM
Sophistication
Moderate
Exploitation Confidence
medium
○CISA SSVC: Public PoC
○Public PoC indexed (trickest/cve)
Composite signal derived from CISA KEV, VulnCheck KEV, CISA SSVC, EPSS, Metasploit, Exploit-DB, trickest/cve, Nuclei templates, and inthewild.io exploitation reports.

What is the attack surface?

AVACPRUISCIA
AVNetwork
ACLow
PRLow
UINone
SUnchanged
CHigh
INone
ALow

What should I do?

5 steps
  1. PATCH

    Upgrade vLLM to 0.14.1 or later — this is the authoritative fix.

  2. WORKAROUND (if patch is not immediately possible): Block multimodal URL inputs at the API gateway; reject or pre-validate any user-supplied URL parameters before they reach the MediaConnector.

  3. NETWORK CONTROLS

    Apply strict egress rules on vLLM pods — deny outbound access to RFC-1918 ranges, cloud metadata endpoints (169.254.169.254, fd00:ec2::254), and Kubernetes service CIDR.

  4. DETECTION

    Alert on outbound HTTP/S requests from vLLM pods to internal RFC-1918 addresses; monitor for access to /metadata, /api/v1, or management endpoints from AI inference pods.

  5. AUDIT

    Review who has access to multimodal inference endpoints — treat inference API access as privileged if multimodal is enabled.

What does CISA's SSVC say?

DecisionTrack*
Exploitationpoc
AutomatableNo
Technical Impactpartial

Source: CISA Vulnrichment (SSVC v2.0). Decision based on the CISA Coordinator decision tree.

How is it classified?

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act
Article 15-Accuracy, Robustness and CybersecurityArticle 9-Risk Management System
ISO 42001
A.6.2-Risk Assessment and Treatment for AIA.8.3-AI System SecurityA.9.4-Information security for AI systems
NIST AI RMF
GOVERN 6.2-Policies for Vulnerability ManagementMANAGE 2.2-Mechanisms to respond to AI risksMEASURE 2.6-Information Security and Data Privacy Risk Monitoring
OWASP LLM Top 10
LLM02:2025-Sensitive Information DisclosureLLM06-Excessive AgencyLLM06:2025-Excessive AgencyLLM07-Insecure Plugin Design

Frequently Asked Questions

What is CVE-2026-24779?

vLLM servers processing multimodal inputs are vulnerable to SSRF via a URL parser inconsistency that bypasses host restrictions — any user with inference API access can pivot into your internal network. In Kubernetes-based AI serving stacks (llm-d, KServe), this enables pod-to-pod reconnaissance and service disruption. Patch to vLLM 0.14.1 immediately; if you cannot patch, disable multimodal URL loading at the network perimeter.

Is CVE-2026-24779 actively exploited?

Proof-of-concept exploit code is publicly available for CVE-2026-24779, increasing the risk of exploitation.

How to fix CVE-2026-24779?

1. PATCH: Upgrade vLLM to 0.14.1 or later — this is the authoritative fix. 2. WORKAROUND (if patch is not immediately possible): Block multimodal URL inputs at the API gateway; reject or pre-validate any user-supplied URL parameters before they reach the MediaConnector. 3. NETWORK CONTROLS: Apply strict egress rules on vLLM pods — deny outbound access to RFC-1918 ranges, cloud metadata endpoints (169.254.169.254, fd00:ec2::254), and Kubernetes service CIDR. 4. DETECTION: Alert on outbound HTTP/S requests from vLLM pods to internal RFC-1918 addresses; monitor for access to /metadata, /api/v1, or management endpoints from AI inference pods. 5. AUDIT: Review who has access to multimodal inference endpoints — treat inference API access as privileged if multimodal is enabled.

What systems are affected by CVE-2026-24779?

This vulnerability affects the following AI/ML architecture patterns: model serving, multimodal AI pipelines, containerized LLM deployments, LLM inference endpoints, AI agent frameworks using vLLM as backend.

What is the CVSS score for CVE-2026-24779?

CVE-2026-24779 has a CVSS v3.1 base score of 7.1 (HIGH). The EPSS exploitation probability is 0.58%.

What is the AI security impact?

Affected AI Architectures

model servingmultimodal AI pipelinescontainerized LLM deploymentsLLM inference endpointsAI agent frameworks using vLLM as backend

MITRE ATLAS Techniques

AML.T0006Active Scanning
AML.T0025Exfiltration via Cyber Means
AML.T0029Denial of AI Service
AML.T0040AI Model Inference API Access
AML.T0049Exploit Public-Facing Application
AML.T0075Cloud Service Discovery

Compliance Controls Affected

EU AI Act: Article 15, Article 9
ISO 42001: A.6.2, A.8.3, A.9.4
NIST AI RMF: GOVERN 6.2, MANAGE 2.2, MEASURE 2.6
OWASP LLM Top 10: LLM02:2025, LLM06, LLM06:2025, LLM07

What are the technical details?

Original Advisory

vLLM is an inference and serving engine for large language models (LLMs). Prior to version 0.14.1, a Server-Side Request Forgery (SSRF) vulnerability exists in the `MediaConnector` class within the vLLM project's multimodal feature set. The load_from_url and load_from_url_async methods obtain and process media from URLs provided by users, using different Python parsing libraries when restricting the target host. These two parsing libraries have different interpretations of backslashes, which allows the host name restriction to be bypassed. This allows an attacker to coerce the vLLM server into making arbitrary requests to internal network resources. This vulnerability is particularly critical in containerized environments like `llm-d`, where a compromised vLLM pod could be used to scan the internal network, interact with other pods, and potentially cause denial of service or access sensitive data. For example, an attacker could make the vLLM pod send malicious requests to an internal `llm-d` management endpoint, leading to system instability by falsely reporting metrics like the KV cache state. Version 0.14.1 contains a patch for the issue.

Exploitation Scenario

An attacker with standard user-level access to a multi-tenant vLLM deployment crafts an image URL that appears to target an allowed external host (e.g., cdn.allowed-host.com) but uses a backslash in the URL path component. The two Python URL parsing libraries used by MediaConnector interpret this differently: the allowlist check sees cdn.allowed-host.com, while the actual HTTP request resolves to an internal address — e.g., http://169.254.169.254/latest/meta-data/ (AWS IMDS) or http://10.96.0.1/api/v1/secrets (Kubernetes API). In an llm-d cluster, the attacker targets the llm-d management endpoint and submits falsified KV cache metrics, causing the scheduler to misroute requests or trigger unnecessary pod scaling, resulting in service degradation across the inference fleet.

Weaknesses (CWE)

CWE-918 — Server-Side Request Forgery: The web server receives a URL or similar request from an upstream component and retrieves the contents of this URL, but it does not sufficiently ensure that the request is being sent to the expected destination.

  • [Implementation, Operation] If there is a known list of applications, IP, hostname, and/or domains to which the product is allowed to communicate, then use allowlists to ensure that outgoing requests are restricted to these known entities. The allowlists could be in the application code or possibly enforced by firewalls or network segmentation [REF-1511].
  • [Implementation, Operation] If the product is expected to be able to connect to arbitrary IP addresses or domains, consider using a denylist that includes private network addresses, internal domain names, etc. The denylist could be in the application code or possibly enforced by firewalls or network segmentation [REF-1511].

Source: MITRE CWE corpus.

CVSS Vector

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

Timeline

Published
January 27, 2026
Last Modified
July 21, 2026
First Seen
January 27, 2026

Related Vulnerabilities