vLLM speculative decoding crashes server via penalty parameters

# vLLM speculative decoding crashes server via penalty parameters

https://gcdradar.vercel.app/incidents/vllm-speculative-decoding-crashes-server-via-penalty-paramet · AIR-2026-0618 · closed 2026-05-06

- Lane: Patch now
- Severity: 2 (Disclosed vulnerability)
- Fix status: Fix status unknown
- First reported: 2026-05-06
- AI role: Attack on AI
- Products: vLLM
- CVE ids: CVE-2026-44223
- CVSS: 6.5

## Summary

A vulnerability in vLLM allows a denial of service attack by sending requests with specific sampling penalty parameters. The issue causes the inference engine to crash deterministically.

## What to do

No action recorded; read the sources below and the vendor's advisory.

## Latest updates

- 2026-05-12: CVE-2026-44223: vLLM is an inference and serving engine for large language models (LLMs). (https://nvd.nist.gov/vuln/detail/CVE-2026-44223)
- 2026-05-06: CVE-2026-44223: vLLM: extract_hidden_states speculative decoding crashes server on any request with penalty parameters (https://github.com/advisories/GHSA-83vm-p52w-f9pw)

## Sources

- 2026-05-06 GitHub Advisory Database: CVE-2026-44223: vLLM: extract_hidden_states speculative decoding crashes server on any request with penalty parameters https://github.com/advisories/GHSA-83vm-p52w-f9pw
- 2026-05-12 NVD: CVE-2026-44223: vLLM is an inference and serving engine for large language models (LLMs). https://nvd.nist.gov/vuln/detail/CVE-2026-44223

Verify against the vendor's advisory before changing production. Data CC BY-SA 4.0, https://gcdradar.vercel.app.