CVE-2026-100648: vLLM: audio size limit bypass enables resource exhaustion

MEDIUM
Published September 26, 2026
CISO Take

vLLM before 0.29.0 does not enforce the VLLM_MAX_AUDIO_CLIP_FILESIZE_MB limit when decoding audio in multimodal chat requests, so an unauthenticated client can submit oversized audio files and force the server to burn memory and CPU. The impact is availability only (CVSS 5.3, no confidentiality or integrity loss), and the risk is moderate: EPSS is low (0.00351, top 73th percentile), there is no public exploit, no Nuclei template, it is not in CISA KEV, and CISA SSVC says TRACK. The blast radius is still worth noting, since vLLM has 95 downstream dependents and 134 other CVEs on record, so many internal and customer-facing inference stacks may inherit the flaw. Upgrade to vLLM 0.29.0 or later, and until then cap request body sizes at the reverse proxy or gateway and keep multimodal chat endpoints behind authentication and rate limiting.

Sources:NVDEPSSGitHub Advisoryvulncheck.comATLASCISA SSVC

What is the risk?

Medium. The flaw is network-reachable, needs no privileges or user interaction, and has low complexity, but it only affects availability (A:L). Exploitation likelihood is low to moderate: EPSS 0.00351, no public exploit, no scanner template, not in KEV. Exposure is highest for internet-facing or multi-tenant vLLM deployments with audio-capable multimodal models enabled and no gateway-level size limits. A single GPU-backed inference node that is starved of memory can disrupt every tenant sharing it, which raises the practical impact beyond the 'low' CVSS availability score. Note that the source data does not give an explicit affected version range beyond 'before 0.29.0'.

How does the attack unfold?

Initial Access
Attacker reaches an unauthenticated or weakly protected vLLM multimodal chat endpoint over the network.
AML.T0040
Exploitation
Attacker submits oversized audio files that bypass the VLLM_MAX_AUDIO_CLIP_FILESIZE_MB limit, forcing full decoding.
AML.T0034.001
Impact
Memory and CPU exhaustion causes slowdowns or OOM crashes, denying inference service to legitimate users.
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
5.3 / 10
EPSS
0.4%
chance of exploitation in 30 days
Higher than 29% of all CVEs
Exploitation Status
No known exploitation
Sophistication
Trivial

What is the attack surface?

AVACPRUISCIA
AVNetwork
ACLow
PRNone
UINone
SUnchanged
CNone
INone
ALow

What should I do?

1 step
  1. 1) Upgrade vLLM to 0.29.0 or later (patched per the GHSA advisory and VulnCheck advisory). 2) Until patched, enforce a maximum request body size at the reverse proxy, API gateway, or WAF in front of vLLM (e.g., Nginx client_max_body_size, Envoy or gateway limits) sized to your legitimate audio use case. 3) Require authentication and per-client rate limiting on chat completion endpoints; do not expose vLLM directly to the internet. 4) Set container and process memory/CPU limits so a decoding spike cannot take down the host, and configure restart and health checks. 5) If audio input is not needed, disable or avoid serving audio-capable models. 6) Detection: alert on unusually large POST bodies to chat completion endpoints, sudden memory or CPU spikes on inference pods, OOM-kill events, and elevated 5xx or timeout rates from the multimodal chat route. 7) Inventory every service that bundles vLLM, since the 95 dependents may pin older versions.

What does CISA's SSVC say?

DecisionTrack
Exploitationnone
AutomatableYes
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-Procedures to respond to and recover from unknown risks
OWASP LLM Top 10
LLM10-Unbounded Consumption

Frequently Asked Questions

What is CVE-2026-100648?

vLLM before 0.29.0 does not enforce the VLLM_MAX_AUDIO_CLIP_FILESIZE_MB limit when decoding audio in multimodal chat requests, so an unauthenticated client can submit oversized audio files and force the server to burn memory and CPU. The impact is availability only (CVSS 5.3, no confidentiality or integrity loss), and the risk is moderate: EPSS is low (0.00351, top 73th percentile), there is no public exploit, no Nuclei template, it is not in CISA KEV, and CISA SSVC says TRACK. The blast radius is still worth noting, since vLLM has 95 downstream dependents and 134 other CVEs on record, so many internal and customer-facing inference stacks may inherit the flaw. Upgrade to vLLM 0.29.0 or later, and until then cap request body sizes at the reverse proxy or gateway and keep multimodal chat endpoints behind authentication and rate limiting.

Is CVE-2026-100648 actively exploited?

No confirmed active exploitation of CVE-2026-100648 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-100648?

1) Upgrade vLLM to 0.29.0 or later (patched per the GHSA advisory and VulnCheck advisory). 2) Until patched, enforce a maximum request body size at the reverse proxy, API gateway, or WAF in front of vLLM (e.g., Nginx client_max_body_size, Envoy or gateway limits) sized to your legitimate audio use case. 3) Require authentication and per-client rate limiting on chat completion endpoints; do not expose vLLM directly to the internet. 4) Set container and process memory/CPU limits so a decoding spike cannot take down the host, and configure restart and health checks. 5) If audio input is not needed, disable or avoid serving audio-capable models. 6) Detection: alert on unusually large POST bodies to chat completion endpoints, sudden memory or CPU spikes on inference pods, OOM-kill events, and elevated 5xx or timeout rates from the multimodal chat route. 7) Inventory every service that bundles vLLM, since the 95 dependents may pin older versions.

What systems are affected by CVE-2026-100648?

This vulnerability affects the following AI/ML architecture patterns: model serving, LLM inference APIs, multimodal chat pipelines, voice/audio AI applications.

What is the CVSS score for CVE-2026-100648?

CVE-2026-100648 has a CVSS v3.1 base score of 5.3 (MEDIUM). The EPSS exploitation probability is 0.35%.

What is the AI security impact?

Affected AI Architectures

model servingLLM inference APIsmultimodal chat pipelinesvoice/audio AI applications

MITRE ATLAS Techniques

AML.T0029Denial of AI Service
AML.T0034.001Resource-Intensive Queries
AML.T0040AI Model Inference API Access

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

What are the technical details?

Original Advisory

vllm before 0.29.0 fails to enforce VLLM_MAX_AUDIO_CLIP_FILESIZE_MB limit in multimodal chat audio decoding, allowing unauthenticated clients to bypass file size restrictions. Attackers can submit oversized audio files through chat endpoints to consume excessive memory and CPU resources during decoding.

Exploitation Scenario

An attacker finds an exposed vLLM OpenAI-compatible endpoint serving an audio-capable multimodal model. Without any credentials, they send chat completion requests with very large audio clips (base64-encoded or referenced by URL) that exceed the intended VLLM_MAX_AUDIO_CLIP_FILESIZE_MB limit. Because the limit is not enforced, the server decodes each file in full, consuming large amounts of memory and CPU. By sending several such requests in parallel, the attacker exhausts resources on the inference node, causing OOM kills or severe slowdowns, so legitimate users and downstream applications (voice assistants, transcription features, agent pipelines) see timeouts and failures.

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:N/UI:N/S:U/C:N/I:N/A:L

Timeline

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

Related Vulnerabilities