BlackTreeCVE Intelligence
← Back to the CVE catalogue
Full vulnerability report · 2026
CVE-2026-90554High confidence

vLLM before 0.28.0 Denial of Service via audio extraction

vllm-project · vLLM

Official source article: GitHub GHSA-936P-M5PV-VVJF ↗. Check the applicable product and release in the original source.

6.9MediumCVSS 4.0
Recommended action
Patch only the product branches with a verified fix

Medium technical severity with no CISA KEV confirmation; remediate through the normal risk-based patch cycle unless local exposure raises the priority. Verified remediation exists for at least one product or source, but 27 structured product or package states remain unresolved. Apply remediation only to the exact product branch confirmed by its source.

Fix availability varies by product
R
Operational reassessment

Published severity in operational context

Open reassessment dashboard →
Published severityMediumOperational priority:Medium, unchanged from published severity.unchanged

Evidence used

  • No CISA KEV confirmation is currently recorded.
  • EPSS is 0.20% for the current model date.

Compensating controls

  • Validate the affected product branch and deploy the verified fixed release.
  • Restrict local access and enforce least privilege on affected hosts.
  • Monitor vendor guidance and exploitation sources for a material change.

Verification

  1. Confirm that the asset runs vllm-project vLLM and falls inside the recorded affected range.
  2. Verify the installed build against the product-specific fixed version after deployment.
  3. Validate exposure, authentication requirements and compensating controls in the actual environment.
  4. Reopen this reassessment when CVSS, KEV, EPSS, exploit evidence or remediation changes.
Mitigation target: As exposure requiresRemediation target: Within 365 days

This automated reassessment organises public evidence. It does not know asset exposure, business impact or control effectiveness and does not replace CVSS or a human risk decision.

Cross-source reconciliation

Remediation availability differs by product scope

Verified remediation exists for at least one product or source, but 27 structured product or package states remain unresolved. Apply remediation only to the exact product branch confirmed by its source.

Direct vendor intelligence

Authoritative vendor CSAF and VEX advisories

Structured product status and remediation from the issuing vendor. Product-state explanations are always visible; large lists can be searched or downloaded.

1 current
CVE-2026-90554 · CSAF 2.0 · revision 3 · finalRed Hat Product Securityvllm: vLLM: Denial of Service via video audio extraction
27 known affected

The vendor explicitly identifies these products as affected by this CVE.

  • rhaii/vllm-cpu-rhel9 as a component of Red Hat AI Inference Server
  • rhaii/vllm-cuda-rhel9 as a component of Red Hat AI Inference Server
  • rhaii/vllm-gaudi-rhel9 as a component of Red Hat AI Inference Server
  • rhaii/vllm-neuron-rhel9 as a component of Red Hat AI Inference Server
  • rhaii/vllm-rocm-rhel9 as a component of Red Hat AI Inference Server
  • rhaii/vllm-spyre-rhel9 as a component of Red Hat AI Inference Server
  • rhaii/vllm-tpu-rhel9 as a component of Red Hat AI Inference Server
  • rhaiis/vllm-cpu-rhel9 as a component of Red Hat AI Inference Server
  • rhaiis/vllm-cuda-rhel9 as a component of Red Hat AI Inference Server
  • rhaiis/vllm-neuron-rhel9 as a component of Red Hat AI Inference Server
  • rhaiis/vllm-rocm-rhel9 as a component of Red Hat AI Inference Server
  • rhaiis/vllm-spyre-rhel9 as a component of Red Hat AI Inference Server
Summary
A flaw was found in vLLM. This vulnerability allows a remote attacker to cause a denial of service by providing a specially crafted, highly compressed video as multimodal input when the NanoNemotronVL model is configured to use audio from video. The lack of proper size or duration limits during audio extraction from the video input can force the server to allocate excessive memory, leading to system instability and preventing legitimate users from accessing the service.
Remediation
To mitigate this issue, disable the `use_audio_in_video` feature if it is not required for your deployment. This prevents the vulnerable audio extraction code path from being exercised. If disabling this feature is not feasible, mitigation for this issue is either not available or the currently available options do not meet the Red Hat Product Security criteria comprising ease of use and deployment, applicability to widespread installation base, or stability.
Optional official sources

National CERT insights
?CERT means Computer Emergency Response Team; CSIRT is the closely related term Computer Security Incident Response Team.

Choose official national sources for this report. Each advisory shows its original language. Your selection is remembered on this device and included in shared links.

Official European source

ENISA European Vulnerability Database

Official EUVD identifiers, advisory evidence and known-exploited context. Missing fields are not treated as evidence of low risk.

1 current
ENISA EUVD identifier

EUVD-2026-76872

No EUVD known-exploited evidence

vLLM versions >=0.10.2 and <0.28.0 do not apply any audio decode-size or duration limit when extracting audio from video input for NanoNemotronVL models. In nano_nemotron_vl.py, _extract_audio_from_videos calls load_audio_pyav(BytesIO(video_bytes)) without the max_duration_s or max_decode_bytes parameters, so neither VLLM_MAX_AUDIO_DECODE_DURATION_S nor VLLM_MAX_AUDIO_DECODE_BYTES is enforced (unlike the direct audio upload path in AudioMediaIO). When a NanoNemotronVL model is served with use_audio_in_video=True, an attacker who supplies a small, highly compressed video as multimodal input can force the server to allocate gigabytes of memory during audio decoding, resulting in a denial of service. Fixed in vLLM 0.28.0.

EUVD state
Present in the current official mapping
Known exploitation
Not present in the current ENISA EUVD known-exploited dataset. This is not proof of no exploitation.
ENISA score
6.9 · CVSS 4.0
Advisory evidence
No linked advisory details stored yet
Recommended actionPatch only the product branches with a verified fix

Medium technical severity with no CISA KEV confirmation; remediate through the normal risk-based patch cycle unless local exposure raises the priority. Verified remediation exists for at least one product or source, but 27 structured product or package states remain unresolved. Apply remediation only to the exact product branch confirmed by its source.

Fix availability varies by product
01

What, why and how

vLLM versions >=0.10.2 and <0.28.0 do not apply any audio decode-size or duration limit when extracting audio from video input for NanoNemotronVL models. In nano_nemotron_vl.py, _extract_audio_from_videos calls load_audio_pyav(BytesIO(video_bytes)) without the max_duration_s or max_decode_bytes parameters, so neither VLLM_MAX_AUDIO_DECODE_DURATION_S nor VLLM_MAX_AUDIO_DECODE_BYTES is enforced (unlike the direct audio upload path in AudioMediaIO). When a NanoNemotronVL model is served with use_audio_in_video=True, an attacker who supplies a small, highly compressed video as multimodal input can force the server to allocate gigabytes of memory during audio decoding, resulting in a denial of service. Fixed in vLLM 0.28.0.

What

vLLM versions >=0.10.2 and <0.28.0 do not apply any audio decode-size or duration limit when extracting audio from video input for NanoNemotronVL models. In nano_nemotron_vl.py, _extract_audio_from_videos calls load_audio_pyav(BytesIO(video_bytes)) without the max_duration_s or max_decode_bytes parameters, so neither VLLM_MAX_AUDIO_DECODE_DURATION_S nor VLLM_MAX_AUDIO_DECODE_BYTES is enforced (unlike the direct audio upload path in AudioMediaIO). When a NanoNemotronVL model is served with use_audio_in_video=True, an attacker who supplies a small, highly compressed video as multimodal input can force the server to allocate gigabytes of memory during audio decoding, resulting in a denial of service. Fixed in vLLM 0.28.0.

Why

The product does not properly control the allocation and maintenance of a limited resource.

How

An attacker operating through local access may attempt exploitation without authentication or user interaction. If successful, the issue may disrupt the affected service.

What

vLLM versions >=0.10.2 and <0.28.0 do not apply any audio decode-size or duration limit when extracting audio from video input for NanoNemotronVL models. In nano_nemotron_vl.py, _extract_audio_from_videos calls load_audio_pyav(BytesIO(video_bytes)) without the max_duration_s or max_decode_bytes parameters, so neither VLLM_MAX_AUDIO_DECODE_DURATION_S nor VLLM_MAX_AUDIO_DECODE_BYTES is enforced (unlike the direct audio upload path in AudioMediaIO). When a NanoNemotronVL model is served with use_audio_in_video=True, an attacker who supplies a small, highly compressed video as multimodal input can force the server to allocate gigabytes of memory during audio decoding, resulting in a denial of service. Fixed in vLLM 0.28.0.

Why

The product does not properly control the allocation and maintenance of a limited resource.

How

An attacker operating through local access may attempt exploitation without authentication or user interaction. If successful, the issue may disrupt the affected service.

02

Exploit reality and attack path

CVSS severity, EPSS forecast probability, public exploit material and CISA-confirmed exploitation are separate signals.

Observed exploitation
?Confirmed exploitation and public exploit material are separate signals. Attacks can occur without public proof-of-concept or exploit code.
No confirmed evidence

No CISA KEV match was present at the last successful refresh. This means no confirmation from that source, not proof of no exploitation.

Public PoC / exploit material
?Confirmed exploitation and public exploit material are separate signals. Attacks can occur without public proof-of-concept or exploit code.
None recorded

No exploit-tagged reference or CISA SSVC proof-of-concept state is currently recorded. Research may still exist outside the structured feeds.

Likely attack path
local access → Uncontrolled Resource Consumption → disrupt the affected service
Attack surface
Local
Privileges required
None: unauthenticated exploitation is possible
User interaction
None
Attack complexity
Low: no specialised conditions are recorded
Security boundary
Not a CVSS 4.0 base metric
Weakness
?CWE means Common Weakness Enumeration: a standard category for the underlying weakness.
CWE-400 ↗

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

CVSS vector
?CVSS means Common Vulnerability Scoring System. The vector records the metric values used to calculate technical severity.
CVSS:4.0/AV:L/AC:L/AT:N/PR:N/UI:N/VC:N/VI:N/VA:H/SC:N/SI:N/SA:N

Common Vulnerability Scoring System 4.0: the compact vector below is decoded into plain language.

AVLocalAttack vector: The attacker needs local access to the vulnerable system.ACLowAttack complexity: No specialised conditions are required beyond attacker-controlled input.ATNoneAttack requirements: No additional deployment or execution condition is required.PRNonePrivileges required: The attacker does not need an account or existing privileges.UINoneUser interaction: No action by another user is required.VCNoneVulnerable-system confidentiality: No direct loss is represented by this metric.VINoneVulnerable-system integrity: No direct loss is represented by this metric.VAHighVulnerable-system availability: A successful attack can cause a major loss.SCNoneSubsequent-system confidentiality: No direct loss is represented by this metric.SINoneSubsequent-system integrity: No direct loss is represented by this metric.SANoneSubsequent-system availability: No direct loss is represented by this metric.
Post-exploitation / living off the land
The issue can support a local privilege or sandbox boundary transition; normal system utilities may then be available in the gained context.
UnauthenticatedDenial of serviceCWE-400
A

Official authority intelligence

Only matched European and national findings are included. Language selectors and unavailable sources are omitted.

ENISA EUVD · EUVD-2026-76872Official EUVD mapping

vLLM versions >=0.10.2 and <0.28.0 do not apply any audio decode-size or duration limit when extracting audio from video input for NanoNemotronVL models. In nano_nemotron_vl.py, _extract_audio_from_videos calls load_audio_pyav(BytesIO(video_bytes)) without the max_duration_s or max_decode_bytes parameters, so neither VLLM_MAX_AUDIO_DECODE_DURATION_S nor VLLM_MAX_AUDIO_DECODE_BYTES is enforced (unlike the direct audio upload path in AudioMediaIO). When a NanoNemotronVL model is served with use_audio_in_video=True, an attacker who supplies a small, highly compressed video as multimodal input can force the server to allocate gigabytes of memory during audio decoding, resulting in a denial of service. Fixed in vLLM 0.28.0.

Official EUVD record ↗
Cyber Security Agency of Singapore · English · CSA-SB-20260916Security Bulletin 16 Sep 2026

The Cyber Security Agency of Singapore included this CVE in its official Security Bulletin 16 Sep 2026, published on 16 September 2026. Open the linked bulletin for the product, severity and reference information published in that issue.

Official advisory ↗
03

Patch and workaround

Operational remediation based on structured source evidence.

Status
?Patch availability is based on structured fixed-version fields and authoritative update references. If no fix is verified, check the vendor advisory before making a change.
Fix availability varies by product
Affected
Fixed
vLLM: 0.28.0
Action
Use the product-specific evidence above. Patch only products with a verified fixed release, and keep every affected or under-investigation state without a matching fix in the remediation queue.
Workaround
To mitigate this issue, disable the `use_audio_in_video` feature if it is not required for your deployment. This prevents the vulnerable audio extraction code path from being exercised. If disabling this feature is not feasible, mitigation for this issue is either not available or the currently available options do not meet the Red Hat Product Security criteria comprising ease of use and deployment, applicability to widespread installation base, or stability.
Official vendor articles1 stored document

Linked articles are downloaded and versioned as source evidence. A CVE mention or approved update relationship does not, by itself, verify a fix for every product branch.

04

Evidence and provenance

Published 12 Sept 2026 · Last source change 14 Sept 2026, 18:16 UTC · CWE-400 · Uncontrolled Resource Consumption

CVE recordCVE.org · 5.2
CVSS sourceCNA
EPSS source
?The date BlackTree first stored a score for this CVE from the daily FIRST EPSS feed.
FIRST · tracked since 2026-09-14
European sourceENISA EUVD · EUVD-2026-76872
Product sourceVendor CSAF · Red Hat Product Security
Remediation sourceVendor CSAF · Red Hat Product Security
CWE sourceCNA
NVD statusNVD enriched

Core structured fields are present and their contributing authorities are shown above.

Material change intelligence

What changed after publication

View recent updates ↗
  1. Vendor guidanceAuthoritative vendor guidance changed from remediation: access.redhat.com/CVE-2026-90554 to remediation: access.redhat.com/CVE-2026-90554.
    Before
    remediation: access.redhat.com/CVE-2026-90554
    After
    remediation: access.redhat.com/CVE-2026-90554
    Red Hat Product Security ↗
  2. Remediation statusRemediation status changed from Awaiting fix to Mitigation available.
    Before
    Awaiting fix
    After
    Mitigation available
    Red Hat Product Security ↗
  3. Vendor guidanceAuthoritative vendor guidance changed: added vendor advisory: github.com/GHSA-936p-m5pv-vvjf.
    Before
    remediation: github.com/GHSA-936p-m5pv-vvjf
    After
    remediation: github.com/GHSA-936p-m5pv-vvjf · vendor advisory: github.com/GHSA-936p-m5pv-vvjf
    disclosure@vulncheck.com ↗
  4. ENISA EUVD mappingEUVD-2026-76872 was added to the official ENISA EUVD mapping for this CVE.
    Before
    not recorded
    After
    {"euvdId":"EUVD-2026-76872"}
    ENISA EUVD ↗
  5. Catalogue recordCVE added to the BlackTree catalogue.
    CNA ↗
Material fields only · duplicate refreshes suppressed · history retained for the configured operational retention period
CVE published
Fixed release or patch reference recorded
ENISA EUVD · EUVD-2026-76872↗Vendor CSAF · Red Hat Product Security↗CVE authority or vendor reference · GitHub GHSA-936P-M5PV-VVJF↗Official CVE record (CNA and ADP data)↗NVD record↗
Technical terms and abbreviations used in this report
CVE
Common Vulnerabilities and Exposures: the public identifier for one disclosed vulnerability.
CVSS
Common Vulnerability Scoring System: a technical severity framework; it is not patching priority by itself.
EPSS
Exploit Prediction Scoring System: FIRST's estimate of the probability that exploitation activity will be observed in the next 30 days; it is a forecast, not confirmation.
CWE
Common Weakness Enumeration: the standard category describing the underlying software or hardware weakness.
CNA
CVE Numbering Authority: an organisation authorised to assign and publish CVE records.
CISA ADP
Cybersecurity and Infrastructure Security Agency Authorized Data Publisher: structured enrichment added to a CVE record.
NVD
National Vulnerability Database: NIST's enrichment service for CVE records.
CERT / CSIRT
A computer security incident response team that publishes warnings or coordinates incident response.
PoC
Proof of concept: public material that demonstrates or helps reproduce exploitation.
CSAF
Common Security Advisory Framework: a machine-readable format for security advisories.
LoTL
Living off the land: abuse of legitimate tools or system functions during an attack.
Free version - for non-commercial use only.CVE-2026-90554 · cve.blacktree.nl