vLLM before 0.29.0 Cross-Request Logits Corruption via allowed_token_ids
13Vexday Risk Score
No sign of exploitation. No public exploitation artifact known so far.
ssvc Trackcvss 6.3epss 0.2%
exploitation probability
0.2%top 84% of all CVEs
observed exploitation
nono source reports it
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:4.0/AV:N/AC:L/AT:P/PR:N/UI:N/VC:N/VI:L/VA:N/SC:N/SI:N/SA:N
Affected products
vllm-project · vllmReferences
https://github.com/vllm-project/vllmhttps://github.com/vllm-project/vllm/blob/v0.28.0/vllm/sampling_params.py#L881-L903https://github.com/vllm-project/vllm/blob/v0.28.0/vllm/v1/worker/gpu/sample/logit_bias.py#L179-L191https://github.com/vllm-project/vllm/commit/5b0e5b69ac1a3884a6479c9537789c95263cc804https://github.com/vllm-project/vllm/pull/49080https://www.vulncheck.com/advisories/vllm-before-0.29.0-cross-request-logits-corruption-via-allowed-token-ids