Exposure of Data Element to Wrong Session in vLLM - #VU165865
Published: November 11, 2026
Vulnerability details
The vulnerability allows a remote user to disrupt another concurrent request's generation by forcing the configured reasoning-end sequence, such as </think>, into its output.
The vulnerability exists due to failure to remove source entries in the V1 model runner's ThinkingBudgetStateHolder.sync_batch() when swapping a budgeted row with an unbudgeted row, aliasing mutable state across requests. A remote user can submit a completion request with thinking_token_budget and groom the queue to trigger a mixed batch swap on a reasoning-enabled server to force the configured reasoning-end sequence into a victim's output.