Advisory

PYSEC-2026-3998

Also known as: CVE-2026-93840

Summary

The record has no summary.

Details

vLLM before 0.29.0 validates allowed_token_ids against tokenizer length instead of model output logits width in SamplingParams._validate_allowed_token_ids(). Attackers can supply token IDs above the output vocabulary that pass validation, causing LogitBiasState to corrupt GPU logits state and allow concurrent requests to sample tokens outside their allowlists.

Severity

Medium Basis: CVSS v3 score 5.3, vector CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:L/A:N

Dates

Published
2026-09-18 20:17 UTC
Modified
2026-09-29 09:00 UTC

Affected packages and fixes

vllm

PyPI · All systems

Affected

  • introduced 0, fixed 0.29.0 (ecosystem)

Fixed versions

0.29.0

Fix

  • Fixed version 0.29.0
    pip install --upgrade vllm==0.29.0

Fix commands are the export's own generic commands. Ojuju never runs them. Check them against your setup first.

Check your own computer: guard packages check

View this record on OSV.dev

Source

Source: PyPA Advisory Database · CC BY 4.0

Sources & licences