# CVE-2026-115913: vLLM Phi-4 multimodal audio encoder GPU memory exhaustion

Source: https://junglewise.ai/threats/cve-2026-115913-vllm-is-an-inference-and-serving-engine-for-large-language  
Publisher: Junglewise Threat Intelligence, https://junglewise.ai/threats  
Published: 2026-10-26. Updated: 2026-10-26  
License: CC BY 4.0 with a link back

CVE-2026-115913 · Severity: medium · CVSS 6.5 · Published 2026-10-26

vLLM is an inference engine for large language models that processes audio and other multimodal inputs. A remote attacker can send a single legitimate long audio file that causes the Phi-4 audio encoder to retain full attention matrices in GPU memory, exhausting GPU resources and crashing the service for all other users.

Technologies: [vllm (PyPI)](https://junglewise.ai/threats/technologies/pypi-vllm), [vLLM Project vllm](https://junglewise.ai/threats/technologies/vllm-project-vllm). Vendors: [PyPI](https://junglewise.ai/threats/vendors/pypi), [vLLM Project](https://junglewise.ai/threats/vendors/vllm-project).

## Quick answers

**Is CVE-2026-115913 exploited in the wild?** No. Junglewise Threat Intelligence has no report of CVE-2026-115913 being exploited in the wild.

**How severe is CVE-2026-115913?** Medium severity, with a CVSS score of 6.5, published 26 October 2026.

**Which products and versions are affected?** vllm (PyPI) < 0.32.0, ECOSYSTEM:introduced=0,fixed=0.32.0; vLLM Project vllm before 0.32.0.

**Is there a fix for CVE-2026-115913?** Yes. vllm (PyPI): upgrade to 0.32.0 or later. Until then: Disable the eager attention path or explicitly configure the Phi-4 audio encoder to use PyTorch scaled-dot-product attention instead of eager attention evaluation. Junglewise finds every place you run it, drafts the upgrade and applies it once you approve.

**How is CVE-2026-115913 attacked?** It is attacked over the network, in a software library.

**Am I affected by CVE-2026-115913?** Check whether you run vllm (PyPI) or vLLM Project vllm at an affected version. Junglewise does that check for you around the clock: connect your code, cloud and domains, and it matches new vulnerabilities like this one against what you actually run and tells you when one reaches you.

## How to fix

- vllm (PyPI): upgrade to 0.32.0 or later

  ```sh
  pip install "vllm>=0.32.0"
  ```


Disable the eager attention path or explicitly configure the Phi-4 audio encoder to use PyTorch scaled-dot-product attention instead of eager attention evaluation.

## Technical details

The Phi-4 multimodal audio encoder in vLLM versions prior to 0.32.0 uses eager attention evaluation by default in MultiHeadedAttention.forward_attention(), which materializes and retains full attention probability matrices in memory across all 24 encoder layers. A remote client can send a maximal-length audio input that causes the conformer encoder to pin multi-gigabyte tensors, triggering GPU out-of-memory and terminating the EngineCore service. The scaled dot-product attention branch does not materialize these matrices; the fix switches the default to use PyTorch's scaled-dot-product attention and removes matrix retention from the eager path.

## Affected products

- vLLM Project vLLM before 0.32.0

## Identifiers

- CVE-2026-115913
- GHSA-8397-8qf6-f2gm

## Timeline

- 2026-10-26: disclosed
- 2026-10-09: patched

## References

- [https://github.com/vllm-project/vllm/commit/3d52578e13454de0bb8352cac988347cdcc89928](https://github.com/vllm-project/vllm/commit/3d52578e13454de0bb8352cac988347cdcc89928)
- [https://github.com/vllm-project/vllm/pull/60614](https://github.com/vllm-project/vllm/pull/60614)

## Related threats

- [vLLM chat template resource exhaustion DoS](https://junglewise.ai/threats/cve-2026-115918-vllm-is-an-inference-and-serving-engine-for-large-language)
- [vLLM: Permissive default CORS origins allow cross-origin requests to the vLLM HTTP server](https://junglewise.ai/threats/cve-2026-115917-vllm-is-an-inference-and-serving-engine-for-large-language)
- [vLLM off-by-one error in tokenizer max_token_id validation](https://junglewise.ai/threats/cve-2026-115916-vllm-is-an-inference-and-serving-engine-for-large-language)
- [vLLM image format validation bypass in media processing](https://junglewise.ai/threats/cve-2026-115915-vllm-is-an-inference-and-serving-engine-for-large-language)
- [vLLM denial of service in DeepSeek-OCR image processing](https://junglewise.ai/threats/cve-2026-115914-vllm-is-an-inference-and-serving-engine-for-large-language)