vLLM: Downmix Implementation Differences as Attack Vectors Against Audio AI Models
vLLM is an inference and serving engine for large language models (LLMs). From version 0.5.5 to before version 0.18.0, Librosa defaults to using numpy.mean for mono downmixing (to_mono), while the international standard ITU-R BS.775-4 specifies a weighted downmixing algorithm. This discrepancy results in inconsistency between audio heard by humans (e.g., through headphones/regular speakers) and audio processed by AI models (Which infra via Librosa, such as vllm, transformer). This issue has been patched in version 0.18.0.
- Published
- Apr 2, 2026
- Updated
- Jul 24, 2026
- EPSS
- 0.5% · 39th percentile
- CWE
- CWE-20
Sign in to follow · You’ll be notified if a followed record enters KEV, gets an exploit or is updated.
Report tools
Action score
28
Monitor
Low priority for now.
- CVSS
- 28 / 40 · 7.1 / 10
- CISA KEV
- 0 / 30 · Not listed
- EPSS
- 0 / 30 · 0.5%
CISA SSVC decision
- Exploitation
- none
- Automatable
- no
- Technical impact
- partial
Vulnrichment: CISA's decision-tree inputs.
CNA vs NVD score
- NVD
- 7.1
- CNA · GitHub_M
- 5.9
- 1.2 point gap
CVSS:3.1/AV:N/AC:H/PR:L/UI:N/S:U/C:N/I:H/A:L
The score given by the assigning authority versus NVD’s independent score. A gap means the severity is contested.
Affected systems
| Vendor | Product | CPE |
|---|---|---|
| vllm | vllm | cpe:2.3:a:vllm:vllm |
Affected versions
NVD version ranges (for catalog products). Add the product to your stack with a version and matching uses these.
- vllm vllm0.5.5 and later · before 0.18.0
Versions reported by the vendor
Affected version ranges reported by the assigning authority (GitHub_M). Independent of NVD's CPE analysis and usually ahead of it.
vllm-project vllm
- >= 0.5.5, < 0.18.0affected
Package-level exposure
OSV and GitHub Advisory data: ecosystem, package and range. SBOM matching uses this table.
| Ecosystem | Package | Affected range | Fix |
|---|---|---|---|
| PyPI | vllm | from 0.5.5 · before 0.18.0 | 0.18.0 |
Same product
vllm: all recordsOther highest-scoring records for the same primary product.
- CVE-2026-22778vLLM leaks a heap address when PIL throws an error42Plan
- CVE-2025-32444vLLM Vulnerable to Remote Code Execution via Mooncake Integration40Plan
- CVE-2026-25960SSRF Protection Bypass in vLLM39Monitor
- CVE-2026-22807vLLM affected by RCE via auto_map dynamic module loading during model initialization39Monitor
- CVE-2025-47277vLLM Allows Remote Code Execution via PyNcclPipe Communication Service39Monitor
- CVE-2024-11041Remote Code Execution in vllm-project/vllm39Monitor
Remediation
Which version to upgrade to
Fix versions compiled from the vendor, package registries and Microsoft. Verify the vendor's note before upgrading.
| Product / package | Fixed version | Source |
|---|---|---|
| vllm-project vllm | 0.18.0 · description | Vendor (CNA) |
| PyPI:vllm | 0.18.0 | Package registry (OSV) |
Exploit status
No known public exploit
No public exploit has been observed yet. That doesn't mean you're safe, only that the bar is a little higher.
Research context
For pentesters and researchers: attack profile, score disagreement, timeline, patch commits, credits, variant and chain candidates, bug bounty scope. All derived from existing data; no exploit code.
Timeline
From publication to today: proof of concept, Metasploit module, CISA KEV and fix record. Dates are as reported by the sources.
No dated events beyond publication.
FIRST EPSS daily score; only changes of 0.01 or more are recorded (step chart).
Patch and commit links
Commit, PR and diff links among the references. A starting point for patch-diffing and variant hunting; fixes, not exploits.
Credits
All researchersFinders, reporters and analysts named in the CNA record. Click a name for that researcher’s other records.
No credits in the CNA record.
Variant candidates
Same product, same weakness class, within 18 months. If the patch missed the root cause, the sibling bug is here.
- CVE-2025-62164133 days apartVLLM deserialization vulnerability leading to DoS and potential RCE35Monitor
- CVE-2026-5634079 days apartvLLM - Denial of Service via Unvalidated Multimodal Embeddings34Monitor
- CVE-2026-5423495 days apartvLLM: Remote DoS in vLLM via Invalid Recovered Token Reinjection30Monitor
- CVE-2025-48944307 days apartvLLM Tool Schema allows DoS via Malformed pattern and type Fields26Monitor
Chain candidates
An authentication bypass and a privilege-requiring bug in the same product, published close together: combined they may become an unauthenticated path.
—
Bug bounty scope
No known public program.
Source: bounty-targets-data (public HackerOne, Bugcrowd, Intigriti, YesWeHack listings).
Technical details
Attack conditions
- Anyone who can reach it over the internet can trigger it.
- A low-privileged account is enough.
- No user action is required.
- No special conditions are required; it is repeatable.
If successful
- Confidentiality
- none
- Integrity
- high · data or configuration can be modified
- Availability
- low · the service can be disrupted
- Attack vector
- Network
- Attack complexity
- Low
- Privileges required
- Low
- User interaction
- None
- Scope
- Unchanged
- Confidentiality impact
- None
- Integrity impact
- High
- Availability impact
- Low
Weakness class (CWE)
CWE-20 · Improper Input ValidationCVSS vector
CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:H/A:L
nvd-primary
Attack context
MITRE CAPEC attack patterns and ATT&CK techniques for this weakness class (CWE). A starting point for detection rules and threat hunting.
Attack patterns (CAPEC)
- Buffer Overflow via Environment VariablesCAPEC-10likelihood: High · High
- Server Side Include (SSI) InjectionCAPEC-101likelihood: High · High
- Cross Zone ScriptingCAPEC-104likelihood: Medium · High
- Command Line Execution through SQL InjectionCAPEC-108likelihood: Low · Very High
- Object Relational Mapping InjectionCAPEC-109likelihood: Low · High
- SQL Injection through SOAP Parameter TamperingCAPEC-110likelihood: High · Very High
- Double EncodingCAPEC-120likelihood: Low · Medium
- Subverting Environment Variable ValuesCAPEC-13likelihood: High · Very High
ATT&CK techniques
- Obfuscated Files or InformationT1027
- Masquerading: Invalid Code SignatureT1036.001
- Steal Web Session CookieT1539
- Subvert Trust Controls: Code SigningT1553.002
- Impair Defenses:Impair Command History LoggingT1562.003
- Hijack Execution Flow:Dynamic Linker HijackingT1574.006
- Hijack Execution Flow:Path Interception by PATH Environment VariableT1574.007
Noroxi analysis
No Noroxi analysis for this record yet
We don't hand-write analysis for all 385,000+ vulnerabilities; that wouldn't be honest. For notable, high-impact vulnerabilities our team writes the mechanism, detection and remediation steps.
We use this product, ask for helpChange log
- Fix✗ → ✓
For records you follow, these changes also arrive as notifications. →
References
- github.com/vllm-project/vllm/commit/c7f98b4d0a63b32ed939e2b6dfaa8a626e9b46c4
- github.com/vllm-project/vllm/pull/37058
- github.com/vllm-project/vllm/releases/tag/v0.18.0
- github.com/vllm-project/vllm/security/advisories/GHSA-6c4r-fmh3-7rh8
Vendor advisories and official records. Exploit/PoC links are deliberately left out.