# vLLM: extract_hidden_states speculative decoding crashes server on any request with penalty parameters

Advisory record · `GHSA-83vm-p52w-f9pw`

- **Identifiers:** `GHSA-83vm-p52w-f9pw` · `CVE-2026-44223` · `PYSEC-2026-145`
- **State:** Confirmed
- **Severity:** CVSS 6.5 Medium
- **Critical:** No
- **Weakness (CWE):** [CWE-131](https://cwe.mitre.org/data/definitions/131.html) · [CWE-704](https://cwe.mitre.org/data/definitions/704.html)
- **Tags:** [Memory corruption](https://agentprime.io/advisories/tag/memory-corruption.md) · [CWE-704](https://agentprime.io/advisories/tag/cwe-704.md)
- **Published:** 6 May 2026
- **Updated:** 10 Sept 2026
- **Reviewed by a person:** No

Published by the GitHub Advisory Database at 6 May 2026, 21:45 UTC. agentprime has not analysed this yet, so whether it affects any app is not known.

## Affected packages

- `vllm` (PyPI): affected `>=0.18.0, <0.20.0`; fixed in `0.20.0`

## Machine-written summary

Summary written by Claude Haiku from the source advisory. Not reviewed by a person.

> vLLM versions 0.18.0 through 0.19.x are affected when the extract_hidden_states speculative decoding method is configured. A single request that sets a sampling penalty parameter (repetition_penalty, frequency_penalty, or presence_penalty) makes the EngineCore process crash with a RuntimeError, which takes the server out of service. Upgrade to vLLM 0.20.0 or later. If upgrading is not possible, avoid extract_hidden_states as the speculative decoding method, or strip penalty parameters from requests at a gateway.

Written 5 Nov 2026.

## Not analysed yet

We haven’t published our analysis of this Advisory yet. Once a person has reviewed it, this page will say what’s vulnerable, what has to be true for your app to be affected, and how to check.

## Severity and exploitation

- 6.5 Medium, CVSS 3.1: `CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H`, from CNA (GitHub_M)

- **CISA KEV:** Not listed.
- **EPSS:** 0.0037, the 30th percentile, on 6 Nov 2026: FIRST’s estimate of the chance of exploitation in the next 30 days.

## Timeline

- 31 Oct 2026, 14:34 UTC: First seen by agentprime
- 31 Oct 2026, 14:34 UTC: [Confirmed by OSV](https://osv.dev/vulnerability/GHSA-83vm-p52w-f9pw)

## Sources

- [GitHub Advisory Database: GHSA-83vm-p52w-f9pw](https://github.com/advisories/GHSA-83vm-p52w-f9pw)
- [OSV: GHSA-83vm-p52w-f9pw](https://osv.dev/vulnerability/GHSA-83vm-p52w-f9pw)
- <https://github.com/vllm-project/vllm/security/advisories/GHSA-83vm-p52w-f9pw>
- <https://nvd.nist.gov/vuln/detail/CVE-2026-44223>
- <https://github.com/vllm-project/vllm/pull/38610>
- <https://github.com/pypa/advisory-database/tree/main/vulns/vllm/PYSEC-2026-145.yaml>
- <https://github.com/vllm-project/vllm>

Facts on this page come from the GitHub Advisory Database (CC-BY 4.0), OSV and the other sources listed, as published. We don’t reproduce full advisory descriptions, which can contain exploit code.

## Check whether this affects you

A Match means an installed version is in the affected range above. Most don’t affect the app; agentprime hasn’t analysed this one yet.

- With the agentprime MCP server (<https://api.agentprime.io/mcp>), ask your agent to check the project: `check_lockfile` sends a lockfile, `check_packages` a list of installed packages, and both return the Matches and the upgrades that clear them. `get_advisory` returns this record. Setup: <https://agentprime.io/mcp.md>
- In a browser: <https://agentprime.io/check/> reads your lockfile in the browser, never uploading it.
- Whether it touches your code, with file and line: the Claude Code plugin, <https://agentprime.io/plugin.md>
- As JSON: <https://api.agentprime.io/v1/advisories/GHSA-83vm-p52w-f9pw>, or `POST https://api.agentprime.io/v1/match` with your packages (<https://agentprime.io/api.md>)

---

The web page: <https://agentprime.io/advisories/GHSA-83vm-p52w-f9pw/>. Everything for agents: <https://agentprime.io/llms.txt>.