MEDIUM 5.3 PyPI
CVE-2026-93840
PYSEC-2026-3998 · CVE-2026-93840
Published · Modified
AI SAST
Find this class of vulnerability in your own code
Corgea's AI-native static analysis detects vulnerabilities like this one across your repositories, ranks them by exploitability, and returns review-ready fixes.
Description
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.
References
- WEB https://github.com/vllm-project/vllm/blob/v0.28.0/vllm/sampling_params.py#L881-L903
- WEB https://github.com/vllm-project/vllm/blob/v0.28.0/vllm/v1/worker/gpu/sample/logit_bias.py#L179-L191
- ADVISORY https://www.vulncheck.com/advisories/vllm-before-0.29.0-cross-request-logits-corruption-via-allowed-token-ids
- FIX https://github.com/vllm-project/vllm/commit/5b0e5b69ac1a3884a6479c9537789c95263cc804
- FIX https://github.com/vllm-project/vllm/pull/49080
- PACKAGE https://github.com/vllm-project/vllm
Ready to move
Start Securing
Free, no credit card | First findings in minutes