CVE-2026-100650: vLLM: unbounded media fetch enables pre-inference DoS

MEDIUMPoC AVAILABLECISA: TRACK*
Published September 26, 2026
CISO Take

vLLM through 0.29.0 downloads and decodes remote or inline media (audio, images and other multimodal inputs) in full before enforcing its own size caps and per-prompt item limits. An authenticated user, or anyone at all on the unauthenticated Rust /tokenize route, can exhaust server memory and outbound bandwidth with a single oversized body or a request carrying many media parts. Exploitation likelihood is low for now (EPSS 0.65%, roughly top 50%, no public exploit, no Nuclei template, not in CISA KEV, SSVC TRACK*), but the attack needs no skill and the vLLM package has 95 downstream dependents and 134 other CVEs on record. There is no code execution or data disclosure, so the risk is availability of your GPU inference tier. Upgrade to a release containing the fix (commit 752a3a5, GHSA-p6g9-7v3x-m8mv), and until then keep /tokenize and multimodal endpoints behind an authenticated gateway with request body size limits.

Sources:NVDEPSSGitHub AdvisoryATLASvulncheck.comCISA SSVC

What is the risk?

Medium (CVSS 6.5, A:H only). Attack is network-reachable with low complexity and no user interaction. Chat and batch paths require an API key only when one is configured, so many internal or dev deployments are effectively open. The Rust frontend /tokenize route is unauthenticated by design, which raises exposure for any internet-facing instance. Real-world exploitation signals are weak (EPSS 0.00653, no public PoC, not in KEV, SSVC TRACK*), so priority is moderate, higher for shared multi-tenant or customer-facing inference services where an OOM kill disrupts every tenant.

How does the attack unfold?

Initial Access
Attacker reaches the vLLM API (using any valid API key, or none for the Rust /tokenize route).
AML.T0049
Exploitation
Attacker sends a request with a huge remote media URL, oversized base64 payload or many media parts, which the server fetches and decodes in full before checking limits.
AML.T0040
Impact
Memory and outbound bandwidth are exhausted, the server or batch runner crashes or stalls, and inference becomes unavailable.
AML.T0029

What systems are affected?

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

Do you use vLLM? You're affected.

How severe is it?

CVSS 3.1
6.5 / 10
EPSS
0.8%
chance of exploitation in 30 days
Higher than 57% of all CVEs
Exploitation Status
Exploit Available
Exploitation: MEDIUM
Sophistication
Trivial
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
CNone
INone
AHigh

What should I do?

1 step
  1. 1) Upgrade vLLM to a release that includes the fix (commit 752a3a504485790a2e8491cacbb35c137339ad34; see GHSA-p6g9-7v3x-m8mv) and verify the installed version is later than 0.29.0. 2) Until patched, place a reverse proxy or API gateway in front of vLLM that enforces a maximum request body size and rejects oversized or deeply nested multimodal payloads. 3) Do not expose the Rust frontend /tokenize route or any vLLM endpoint to untrusted networks; require API keys and network allow-lists. 4) Restrict remote media fetching with an egress allow-list for audio_url and similar URLs, and block internal ranges. 5) Set container memory limits and restart policies so a single process failure does not take down the node. 6) Detect abuse by alerting on spikes in request body size, outbound bandwidth from inference hosts, per-request media part counts, and OOM-kill or restart events.

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 cybersecurity
ISO 42001
A.6.2.6-AI system operation and monitoring
NIST AI RMF
MANAGE 2.3-Respond to and recover from unknown or previously unseen risks
OWASP LLM Top 10
LLM10:2025-Unbounded Consumption

Frequently Asked Questions

What is CVE-2026-100650?

vLLM through 0.29.0 downloads and decodes remote or inline media (audio, images and other multimodal inputs) in full before enforcing its own size caps and per-prompt item limits. An authenticated user, or anyone at all on the unauthenticated Rust /tokenize route, can exhaust server memory and outbound bandwidth with a single oversized body or a request carrying many media parts. Exploitation likelihood is low for now (EPSS 0.65%, roughly top 50%, no public exploit, no Nuclei template, not in CISA KEV, SSVC TRACK*), but the attack needs no skill and the vLLM package has 95 downstream dependents and 134 other CVEs on record. There is no code execution or data disclosure, so the risk is availability of your GPU inference tier. Upgrade to a release containing the fix (commit 752a3a5, GHSA-p6g9-7v3x-m8mv), and until then keep /tokenize and multimodal endpoints behind an authenticated gateway with request body size limits.

Is CVE-2026-100650 actively exploited?

Proof-of-concept exploit code is publicly available for CVE-2026-100650, increasing the risk of exploitation.

How to fix CVE-2026-100650?

1) Upgrade vLLM to a release that includes the fix (commit 752a3a504485790a2e8491cacbb35c137339ad34; see GHSA-p6g9-7v3x-m8mv) and verify the installed version is later than 0.29.0. 2) Until patched, place a reverse proxy or API gateway in front of vLLM that enforces a maximum request body size and rejects oversized or deeply nested multimodal payloads. 3) Do not expose the Rust frontend /tokenize route or any vLLM endpoint to untrusted networks; require API keys and network allow-lists. 4) Restrict remote media fetching with an egress allow-list for audio_url and similar URLs, and block internal ranges. 5) Set container memory limits and restart policies so a single process failure does not take down the node. 6) Detect abuse by alerting on spikes in request body size, outbound bandwidth from inference hosts, per-request media part counts, and OOM-kill or restart events.

What systems are affected by CVE-2026-100650?

This vulnerability affects the following AI/ML architecture patterns: model serving, LLM inference APIs, multimodal pipelines, speech/audio batch processing, OpenAI-compatible API gateways.

What is the CVSS score for CVE-2026-100650?

CVE-2026-100650 has a CVSS v3.1 base score of 6.5 (MEDIUM). The EPSS exploitation probability is 0.81%.

What is the AI security impact?

Affected AI Architectures

model servingLLM inference APIsmultimodal pipelinesspeech/audio batch processingOpenAI-compatible API gateways

MITRE ATLAS Techniques

AML.T0029Denial of AI Service
AML.T0034.001Resource-Intensive Queries
AML.T0049Exploit Public-Facing Application

Compliance Controls Affected

EU AI Act: Article 15
ISO 42001: A.6.2.6
NIST AI RMF: MANAGE 2.3
OWASP LLM Top 10: LLM10:2025

What are the technical details?

Original Advisory

vLLM through 0.29.0 fetches and fully materializes remote or inline media before enforcing its documented media controls (the VLLM_MAX_AUDIO_CLIP_FILESIZE_MB compressed-audio size cap, default 25 MB, and the per-modality --limit-mm-per-prompt item limits). Across four ingress paths — the shared media-acquisition layer (HTTPConnection.get_bytes()/async_get_bytes()), the chat completions audio_url/base64 path, the batch speech runner, and the Rust frontend POST /tokenize route — the server reads the entire HTTP response body, base64-decodes the inline payload, or spawns one fetch/decode task per media part, and only then applies the limit (or, on some paths, never applies it). A remote attacker can therefore cause the API server or batch-runner process to allocate memory and consume outbound bandwidth proportional to an attacker-chosen body size or media item count before the request is rejected, resulting in pre-inference memory and bandwidth exhaustion (denial of service). The chat and batch surfaces require an API key when one is configured; the Rust frontend /tokenize route is unauthenticated by design. There is no code execution or data disclosure impact.

Exploitation Scenario

An attacker finds an exposed vLLM deployment, for example a hosted speech or multimodal API. They submit chat completion requests whose audio_url points to an attacker-controlled server that streams a multi-gigabyte response, or whose inline base64 payload is extremely large, or whose prompt lists hundreds of media parts that each spawn a fetch and decode task. The server buffers everything in memory before the size or item limit is checked, so a handful of concurrent requests exhaust RAM and bandwidth, the process is OOM-killed, and the inference service goes down repeatedly. If the Rust frontend is exposed, the same effect is achievable against /tokenize without any credentials.

Weaknesses (CWE)

CWE-400 — Uncontrolled Resource Consumption: The product does not properly control the allocation and maintenance of a limited resource.

  • [Architecture and Design] Design throttling mechanisms into the system architecture. The best protection is to limit the amount of resources that an unauthorized user can cause to be expended. A strong authentication and access control model will help prevent such attacks from occurring in the first place. The login application should be protected against DoS attacks as much as possible. Limiting the database access, perhaps by caching result sets, can help minimize the resources expended. To further limit the potential for a DoS attack, consider tracking the rate of requests received from users and blocking requests that exceed a defined rate threshold.
  • [Architecture and Design] Mitigation of resource exhaustion attacks requires that the target system either: The first of these solutions is an issue in itself though, since it may allow attackers to prevent the use of the system by a particular valid user. If the attacker impersonates the valid user, they may be able to prevent the user from accessing the server in question. The second solution is simply difficult to effectively institute -- and even when properly done, it does not provide a full solution. It simply makes the attack require more resources on the part of the attacker. recognizes the attack and denies that user further access for a given amount of time, or uniformly throttles all requests in order to make it more difficult to consume resources more quickly than they can again be freed.

Source: MITRE CWE corpus.

CVSS Vector

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

Timeline

Published
September 26, 2026
Last Modified
October 6, 2026
First Seen
September 26, 2026

Related Vulnerabilities