CVE-2026-114049

Versions of vLLM prior to 0.31.0 are vulnerable to a cross-request mask leak within the InputBatch.condense() function. This flaw occurs because the application leaves stale allowed_token_ids rows, which can then be applied to subsequent requests. A remote attacker can exploit this by submitting ...

CVE-2026-114049 vLLM before 0.31.0 Stale allowed_token_ids Mask via InputBatch.condense

vLLM before 0.31.0 contains a cross-request mask leak in InputBatch.condense that allows remote attackers to constrain other clients' outputs by leaving stale allowedtokenids rows. Attackers can submit concurrent constrained generate requests, including via unauthenticated /invocations, so recycl...