vLLMPYSEC-2026-3998
vLLM before 0.29.0 validates allowed_token_ids against tokenizer length instead...
Fix: upgrade to 0.29.0 or later
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.
Affected versions
| Package | Affected | Fixed in |
|---|---|---|
| vllm PyPI | < 0.29.0 | 0.29.0 |
Details and references
- CVSS 3.1
- CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:L/A:N
- Severity from
- the CVSS score
- Also known as
- CVE-2026-93840
- github.com/vllm-project/vllm/blob/v0.28.0/vllm/sampling_params.py#L881-L903
- github.com/vllm-project/vllm/blob/v0.28.0/vllm/v1/worker/gpu/sample/logit_bias.py#L179-L191
- www.vulncheck.com/advisories/vllm-before-0.29.0-cross-request-logits-corruption-via-allowed-token-ids
- github.com/vllm-project/vllm/commit/5b0e5b69ac1a3884a6479c9537789c95263cc804
- github.com/vllm-project/vllm/pull/49080
- github.com/vllm-project/vllm