PYSEC-2026-4349: vLLM /v1/responses chat-template trust-policy bypass
In vLLM versions 0.16.0 through before 0.32.0, an authenticated requester can use the /v1/responses endpoint to provide an attacker-controlled Jinja chat template when no server template is configured. This can cause CPU or memory exhaustion. The issue is fixed in version 0.32.0.
CERT24