LiteLLM moved prompt-injection heuristics off the asyncio event loop.

v1.103.0-rc.1 on September 20 includes #41541, fixing a guardrail that blocked the event loop for every request it inspected, plus #41555 which returns 429 when a deployment's max_parallel_requests slots are full instead of queueing. Also JWT key-mapping cache eviction on user, team, org and bulk key deletion, and save_config persisting only the keys a caller changed. A security control that halves your throughput gets turned off, so this is a security fix dressed as a performance fix.