Executive brief
vLLM is an open-source serving engine for large language models. A performance vulnerability in the thinking budget state tracking mechanism causes the system to rescan the entire output history repeatedly, leading to quadratic time complexity that degrades performance under load. An attacker can trigger this inefficiency remotely by sending appropriate inputs, potentially causing service slowdowns or denial of service.
Technical details
The vulnerability is an algorithmic complexity issue (CWE-407) in the thinking_budget_state.py component. When output markers are absent, the state machine rescans the entire output history on each decode step, resulting in O(L²) complexity instead of O(L). This is a remote attack vector with no authentication required—an attacker can craft inference requests that trigger the inefficient code path. The consequence is degraded inference performance and potential service unavailability. A fix has been proposed in pull request #51133 to restore incremental marker search cursors, but it was pending acceptance at the time of advisory publication.
Affected products
- vLLM Project vLLM up to 0.29.0
Timeline
- 2026-09-16: disclosed
- 2026-09-16: advisory: CVE-2026-92365 published