CVE-2026-115913 — vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine
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.
Who exploits it and how urgent it is
A 0–100 score combining confirmed exploitation, ransomware and APT use, public exploits, EPSS and severity.
How the score was calculated
- Technical severity (medium)+3
The score orders the patch queue; it does not replace your asset context (internet exposure, criticality, compensating controls).
Triage
Is it exploited, how likely is exploitation, what does it touch, and how severe do the sources call it.
Remediation
Fixes and mitigations documented by the sources — linked to the original advisory.
No fix documented by the sources we poll. That does not mean none exists: check the vendor advisory.
First 24 hours
Ordered from the record’s own fields — exposure first, because you cannot patch what you have not found. A rule-generated starting point.
- Identify exposed assets running the affected vendor/product/version combinations (Vllm-project Vllm-project/vllm).
- No fix documented in the sources: reduce exposure and monitor the vendor advisory.
- Search available logs for exploit probes matching the vulnerability class (denial of service), errors and authentication anomalies.
Affected scope
Vendor, product and versions as the sources word them.
| Vendor | Product | Versions | Status | Source |
|---|---|---|---|---|
| Vllm-project | Vllm-project/vllm | version < 0.32.0 | Vulnerable | CVE Program (cve.org) e CISA ADP |
Attack characteristics
The CVSS vector, decoded. It describes the attack, not your exposure to it.
Every base score collected
Sources score independently and may disagree; each row says who scored it and under which version.
| Score | Version | Severity | Expl. | Impact | Scored by |
|---|---|---|---|---|---|
| 6.5 | CVSS 3.1 | Medium | 2.8 | 3.6 | security-advisories@github.com via NIST National Vulnerability Database |
| 6.5 | CVSS 3.1 | Medium | — | — | GitHub_M via CVE Program (cve.org) e CISA ADP |
CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H (score shown at the top: security-advisories@github.com) Weakness MITRE
Weakness class according to MITRE’s CWE catalog.
CWE-770 · Allocation of Resources Without Limits or Throttling
Detection
Read off the CVSS vector and the weakness class. Starting points, not tested rules.
- Prioritize edge telemetry for the network-reachable component (proxy, WAF, IDS/IPS).
- Track CPU/memory spikes and availability drops correlated with request patterns.
Timeline
What happened to this CVE, newest first.
- 2026
- Source status changed · Published → Awaiting Analysis
- Record published
- CVSS changed · 6.5
- Source status changed · → Published
- Record created in this database
- CVE published
Source activity
Repeated readings are counted rather than listed.
- 1× Official description updated · Oct 26, 2026 – Oct 26, 2026
References
4 on the record
Elsewhere on this site
Related CVEs (same products)
- Medium CVE-2026-115918 vLLM: Responses API bypasses trust_request_chat_template for chat_template_kwargs 7h
- Medium CVE-2026-115916 vLLM: Out-of-vocabulary prompt token id passes input validation and triggers a CUDA device-side assert (engine crash / DoS) 7h
- Low CVE-2026-115917 vLLM: Permissive default CORS origins allow cross-origin requests to the vLLM HTTP server 7h
- High CVE-2026-115915 vLLM: Untrusted EPS inputs can trigger denial of service or Ghostscript RCE 7h
- Medium CVE-2026-115914 vLLM DeepSeek-OCR image processing vulnerable to denial of service 7h
- Medium CVE-2026-115911 vLLM: DoS via unbounded chat_template Jinja2 rendering exhausts the renderer thread pool 8h
Not in any source we poll
Listed rather than left blank: an empty field and an unmeasured one look identical on screen, and only one is a reason to look elsewhere.
- No FIRST EPSS score.
- No fixed version, patch or mitigation documented.
- No CISA SSVC assessment.
- No confirmed IOCs, IP addresses, domains, file hashes, or malware artifacts supplied.
- No organization-specific asset inventory, compensating-control status, or patch deployment evidence supplied.
- No exploit packet captures, log samples, or incident case IDs supplied.
Answered from this record
Text assembled only from the fields above — no additional claims.
What should defenders know first?
CVE-2026-115913 is "vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine", rated Medium. It affects Vllm-project/vllm (Vllm-project); versions are in the scope table. It is not in CISA KEV, which does not rule out exploitation. CVSS 6.5 describes technical severity. First verify whether exposed or business-critical assets run the affected versions, apply vendor fixes or mitigations, and preserve logs for review.
Comments
Moderated technical discussion. Read the community guidelines.
No comments yet.
Sign in or create an account to comment and suggest corrections.