CVE-2026-115913: vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM…
Noroxi assessment
Priority: Monitor (action score 26/100)
- The sources we track (CISA KEV, Exploit-DB, Metasploit, Nuclei, GitHub) show no public exploit code or exploitation in the wild.
- Probability of exploitation in the next 30 days: 0.32% (FIRST EPSS).
- 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.
- Fixed version: vllm-project vllm 0.32.0.
Low priority for now. Upgrade to vllm-project vllm 0.32.0 or later.
Generated automatically from the record's data: CVSS (NVD and CNA), CISA KEV and SSVC, FIRST EPSS, exploit sources, fixed versions. How we calculate it
Primary sources: NVD · CVE.org · FIRST EPSS
vLLM is an inference and serving engine for large language models. Prior to 0.32.0, the Phi-4 multimodal audio encoder in phi4mm_audio.py and phi4mm_utils.py uses eager attention by default, allowing a remote client to submit one legal long audio clip that causes MultiHeadedAttention.forward_attention to materialize and retain full attention matrices in self.attn. The scaled dot-product attention branch does not materialize these matrices. The resulting quadratic GPU-memory use can cause an out-of-memory condition and terminate EngineCore for all users. This issue is fixed in version 0.32.0.
- Published
- Oct 26, 2026
- Updated
- Nov 2, 2026
- EPSS
- 0.3% · 26th percentile
- CWE
- CWE-770
Report tools
Action score
26
Low priority for now.
- CVSS
- 26 / 40 · 6.5 / 10
- CISA KEV
- 0 / 30 · Not listed
- EPSS
- 0 / 30 · 0.3%
CISA SSVC decision
- Exploitation
- none
- Automatable
- no
- Technical impact
- partial
Vulnrichment: CISA's decision-tree inputs.
CNA vs NVD score
- NVD
- —
- CNA · GitHub_M
- 6.5
- NVD has not scored this yet; the score shown is the CNA’s.
The score given by the assigning authority versus NVD’s independent score. A gap means the severity is contested.
Get told when this record changes
- Not in KEV: no record of exploitation in the wild
- No public exploit code on record
- A fix has been released
If any of these changes, we'll tell you in the daily email digest.
Is your version affected? Check itNoroxi analysis
What does it mean in your environment?
The assessment above is generated automatically from the record's data; it doesn't know your environment. Our team verifies on your systems whether this vulnerability is actually exploitable and which mitigation is enough.
We use this product, ask for helpAffected systems
Is your version affected?
Pick the product and type the version you run. It's compared with this record's version ranges in your browser; the version isn't sent anywhere.
| Vendor | Product | CPE |
|---|---|---|
| vllm | vllm | cpe:2.3:a:vllm:vllm |
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
- < 0.32.0affected
Affected versions
NVD version ranges (for catalog products). Add the product to your stack with a version and matching uses these.
- vllm vllm0.31.0 and later · before 0.32.0
Package-level exposure
OSV and GitHub Advisory data: ecosystem, package and range. SBOM matching uses this table.
| Ecosystem | Package | Affected range | Fix |
|---|---|---|---|
| PyPI | vllm | before 0.32.0 | 0.32.0 |
Same product
vllm: all recordsOther highest-scoring records for the same primary product.
- CVE-2026-22778vLLM leaks a heap address when PIL throws an error60This week
- CVE-2025-32444vLLM Vulnerable to Remote Code Execution via Mooncake Integration40Plan
- CVE-2024-11041Remote Code Execution in vllm-project/vllm40Plan
Published before and after this one in the same product
- ← EarlierCVE-2026-115911Oct 26, 2026 · vLLM: DoS via unbounded chat_template Jinja2 rendering exhausts the renderer thread pool26Monitor
- ← EarlierCVE-2026-115912Oct 26, 2026 · vLLM: Public multimodal SHM handle forgery reaches worker-side `pickle.loads()` with insecure serialization disabled35Monitor
- Later →CVE-2026-115914Oct 26, 2026 · vLLM DeepSeek-OCR image processing vulnerable to denial of service26Monitor
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 | 0.32.0 · description | Vendor (CNA) |
| PyPI:vllm | 0.32.0 | 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.
- Published
- Fix recordedsame day
- TodayNov 14, 2026+18 d
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.
Variant candidates
Same product, same weakness class, within 18 months. If the patch missed the root cause, the sibling bug is here.
- CVE-2026-1161791 days apartvLLM 0.18.0 before 0.31.0 Memory Exhaustion via media_io_kwargs num_frames Override34Monitor
- CVE-2026-1140525 days apartvLLM 0.25.0 before 0.31.0 MOSS-Audio Memory Exhaustion via /tokenize mm_processor_kwargs34Monitor
- CVE-2026-9462435 days apartvLLM through 0.29.0 Denial of Service via Unbounded P2P KV Offloading Sessions34Monitor
- CVE-2026-22773289 days apartvLLM is vulnerable to DoS in Idefics3 vision models via image payload with ambiguous dimensions30Monitor
- CVE-2026-5717340 days apartvLLM: Unauthenticated audio decompression-bomb DoS in /v1/chat/completions26Monitor
- CVE-2026-34755203 days apartvLLM Affected by Denial of Service via Unbounded Frame Count in video/jpeg Base64 Processing26Monitor
Not on record for this CVE: researcher credit · public bounty program.
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
- none
- Availability
- high · the service can be disrupted
- Attack vector
- Network
- Attack complexity
- Low
- Privileges required
- Low
- User interaction
- None
- Scope
- Unchanged
- Confidentiality impact
- None
- Integrity impact
- None
- Availability impact
- High
Weakness class (CWE)
CWE-770 · Allocation of Resources Without Limits or ThrottlingCVSS vector
CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H
nvd-secondary
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)
- FloodingCAPEC-125likelihood: High · Medium
- Excessive AllocationCAPEC-130likelihood: Medium · Medium
- XML Ping of the DeathCAPEC-147likelihood: Low · Medium
- Exponential Data ExpansionCAPEC-197likelihood: High · Medium
- Serialized Data Parameter BlowupCAPEC-229likelihood: High · High
- Serialized Data with Nested PayloadsCAPEC-230likelihood: Medium · High
- Oversized Serialized Data PayloadsCAPEC-231likelihood: Medium · High
- HTTP DoSCAPEC-469Low
ATT&CK techniques
- Network Denial of Service:Direct Network FloodT1498.001
- Network Denial of Service:Reflection AmplificationT1498.002
- Endpoint Denial of ServiceT1499
- Endpoint Denial of Service: OS Exhaustion FloodT1499.001
- Endpoint Denial of Service:Service Exhaustion FloodT1499.002
- Endpoint Denial of Service:Application Exhaustion FloodT1499.003
Change log
- SSVC exploitation— → none
- Exploit maturity— → none
- Fix✗ → ✓
For records you follow, these changes also arrive as notifications. →
References
Show all references (2 more)Show less
Vendor advisories and official records. Exploit/PoC links are deliberately left out.