3.7Low
CVE-2026-93840
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.
What this means for your business
- An attacker can use it remotely, over a network, without a login, and without anyone at your company clicking anything.
- FIRST's prediction model gives it a 0.2% chance of attack attempts being seen in the next 30 days, ranking above 16% of all known flaws.
What to do
- 1Ask your IT team or provider whether any of your systems use the affected product.
- 2If you do, follow the vendor's guidance. No patch reference has been published yet.
Not sure if your company is exposed?
Fastnexa’s certified penetration testers can check whether attackers could use this flaw, or others like it, against your websites, apps and network. The full test is free for our first 10 founding clients until 31 December 2026.
Scoring
- CVSS
- 3.7 (v3.1)
- Vector
CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:N/I:L/A:N- CVSS v4
- 6.3
- Weakness
- CWE-129
- Assigned by
- disclosure@vulncheck.com
Dates
- Published
- 2026-09-18
- Last modified
- 2026-09-21
- Sources
- NVD
References
- https://github.com/vllm-project/vllm
- https://github.com/vllm-project/vllm/blob/v0.28.0/vllm/sampling_params.py#L881-L903
- https://github.com/vllm-project/vllm/blob/v0.28.0/vllm/v1/worker/gpu/sample/logit_bias.py#L179-L191
- https://github.com/vllm-project/vllm/commit/5b0e5b69ac1a3884a6479c9537789c95263cc804
- https://github.com/vllm-project/vllm/pull/49080
- https://www.vulncheck.com/advisories/vllm-before-0.29.0-cross-request-logits-corruption-via-allowed-token-ids