Sensitive Information in Resource Not Removed Before Reuse in vLLM - #VU168131
Published: November 19, 2026
Vulnerability details
The vulnerability allows a remote attacker to receive model responses that reveal or are influenced by token data from another user's request.
The vulnerability exists due to incomplete token-state initialization in the V1 model runner when asynchronous scheduling and pipeline parallelism are both enabled, allowing stale tokens from a reused batch slot to enter the key-value cache. A remote attacker can submit inference requests to an affected deployment to potentially receive a response contaminated by another user's token data.