CVE-2026-93840
vLLM before 0.29.0 Cross-Request Logits Corruption via allowed_token_ids
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.
Attack profile
How easy it is to exploit
- Attack vector
- Network
- Complexity
- Low
- Privileges needed
- None
- User interaction
- None
What an attacker can affect
- Confidentiality
- None
- Integrity
- Low
- Availability
- None
Exploit likelihood
What each source reports
CVE List Score used
vLLM before 0.29.0 Cross-Request Logits Corruption via allowed_token_ids
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.
- Record
CVE-2026-93840- CVSS 4.0
- 6.3 Medium
CVSS:4.0/AV:N/AC:L/AT:P/PR:N/UI:N/VC:N/VI:L/VA:N/SC:N/SI:N/SA:N - CVSS 3.1
- 3.7 Low
CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:N/I:L/A:N - Weakness
- CWE-129
- Updated
- 2026-11-02
EUVDEUVD-2026-83284
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.
- Record
EUVD-2026-83284- CVSS 4.0
- 6.3 Medium
CVSS:4.0/AV:N/AC:L/AT:P/PR:N/UI:N/VC:N/VI:L/VA:N/SC:N/SI:N/SA:N - Updated
- 2026-11-02
GitHub AdvisoriesGHSA-97fp-7cr2-23qc
vLLM before 0.29.0 validates allowed_token_ids against tokenizer length instead of model output...
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.
- Record
GHSA-97fp-7cr2-23qc- CVSS 4.0
- 6.3 Medium
CVSS:4.0/AV:N/AC:L/AT:P/PR:N/UI:N/VC:N/VI:L/VA:N/SC:N/SI:N/SA:N/E:X/CR:X/IR:X/AR:X/MAV:X/MAC:X/MAT:X/MPR:X/MUI:X/MVC:X/MVI:X/MVA:X/MSC:X/MSI:X/MSA:X/S:X/AU:X/R:X/V:X/RE:X/U:X - CVSS 3.1
- 3.7 Low
CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:N/I:L/A:N - Weakness
- CWE-129
- Updated
- 2026-11-02
NVD
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.
- Record
CVE-2026-93840- CVSS 4.0
- 6.3 Medium
CVSS:4.0/AV:N/AC:L/AT:P/PR:N/UI:N/VC:N/VI:L/VA:N/SC:N/SI:N/SA:N/E:X/CR:X/IR:X/AR:X/MAV:X/MAC:X/MAT:X/MPR:X/MUI:X/MVC:X/MVI:X/MVA:X/MSC:X/MSI:X/MSA:X/S:X/AU:X/R:X/V:X/RE:X/U:X - CVSS 3.1
- 3.7 Low
CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:N/I:L/A:N - CVSS 3.1
- 5.3 Medium
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:L/A:N - Weakness
- CWE-129
- Updated
- 2026-11-02
OSV
vLLM before 0.29.0 Cross-Request Logits Corruption via allowed_token_ids
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.
- Record
CVE-2026-93840- CVSS 4.0
- 6.3 Medium
CVSS:4.0/AV:N/AC:L/AT:P/PR:N/UI:N/VC:N/VI:L/VA:N/SC:N/SI:N/SA:N - Weakness
- CWE-129
- Updated
- 2026-11-02
OSVPYSEC-2026-3998
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.
- Record
PYSEC-2026-3998- CVSS 3.1
- 5.3 Medium
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:L/A:N - Updated
- 2026-11-02
Affected products
Affected packages
| Package | Affected | Fixed |
|---|---|---|
vllmPyPI |
before 0.29.0, plus listed versions 0.0.10.1.00.1.10.1.20.1.3 Show 45 moreShow less0.1.40.1.50.1.60.1.70.10.00.10.10.10.1.10.10.20.11.00.11.10.11.20.12.00.13.00.14.00.14.10.15.00.15.10.16.00.17.00.17.10.18.00.18.10.19.00.19.10.2.00.2.10.2.1.post10.2.20.2.30.2.40.2.50.2.60.2.70.20.00.20.10.20.20.21.00.22.00.22.10.23.00.24.00.25.00.25.10.26.00.27.0 |
0.29.0 |
History
| Recorded | Changed | Sources |
|---|---|---|
| first indexed | CVE List, EUVD, GitHub Advisories, NVD, OSV |
Dates are when Every CVE recorded each change. Upstream dates may differ.
In short
CVE-2026-93840 is a medium vulnerability (CVSS 6.3) in Vllm, published September 18, 2026. No source lists public exploit code. EPSS puts the chance of exploitation in the next 30 days at 0.25%, the 15th percentile. Fixed in vllm 0.29.0.
References
github.com/vllm-project/vllm/pull/49080
vulncheck.com/advisories/vllm-before-0.29.0-cross-request-logits-corruption-via-allowed-token-ids
nvd.nist.gov/vuln/detail/CVE-2026-93840
Left out here: 46 versions. The JSON record has them.