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 Parameters
Attack Vector
Network
Can be exploited remotely
Attack Complexity
Low
Easy to exploit
Attack Requirements
Present
Additional conditions required
Privileges Required
None
No privileges needed
User Interaction
None
No user interaction needed
Impact Assessment
Confidentiality
None
No data leak
Integrity
Low
Partial data modification
Availability
None
No disruption
CVSS Vector v4.0
Weakness Type (CWE)
References 6
https://github.com/vllm-project/vllm
disclosure@vulncheck.com
https://github.com/vllm-project/vllm/blob/v0.28.0/vllm/sampling_params.py#L881-…
disclosure@vulncheck.com
https://github.com/vllm-project/vllm/blob/v0.28.0/vllm/v1/worker/gpu/sample/log…
disclosure@vulncheck.com
https://github.com/vllm-project/vllm/commit/5b0e5b69ac1a3884a6479c9537789c95263…
disclosure@vulncheck.com
https://github.com/vllm-project/vllm/pull/49080
disclosure@vulncheck.com
https://www.vulncheck.com/advisories/vllm-before-0.29.0-cross-request-logits-co…
disclosure@vulncheck.com