6.3

CVE-2026-93840: vLLM before 0.29.0 Cross-Request Logits Corruption via allowed_token_ids

Analyzed
AliasesPYSEC-2026-3998
Published 18-09-2026Last modified 24-09-2026

Vulnerability summary

CVE-2026-93840 is a medium (CVSS 6.3) Improper Validation of Array Index vulnerability in vllm-project vllm. EPSS estimates a 0.25% chance of exploitation in the next 30 days, higher than 14.9% of scored vulnerabilities.

Overall risk
25/100
0100
Medium · default modelHow is this calculated?
CVSS score
6.3/10
010
Medium · v4.0 (CVE.org)
EPSS score
0.25%
0%100%
Low · 0.00 pp in 30 days
KEVNot listedNot listed by any tracked catalog.
RansomwareNot flaggedNo ransomware use recorded.
Public exploits0No public exploits recorded

Timeline

  1. 18-09-2026
    PublishedVulnerability first disclosed.
  2. 24-09-2026
    Last modifiedVulnerability information updated.

Description

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.

CVSS metrics

5 scores · 3 sources · 2 versions

EPSS trend

Probability

0.25%

Chance of exploitation in the next 30 days

Percentile

15%

Versus all other CVEs

30-day change

No change 0.00 pp

From 0.25% on 22-09-2026

22-09-202622-10-2026

Known exploited

Compare KEV catalogs
CISA KEV

Not in the CISA KEV catalog.

ENISA CNW

Not in the ENISA CNW catalog.

CIRCL

Not in the CIRCL catalog.

Public exploits

0 found

No public exploits indexed.

Weaknesses

CWE · root cause

Affected systems

PyPI1 product
  • vllm< 0.29.0
Vllm1 product
Vllm-project1 product

3 vendors · 3 products

References

6 links