CVE-2026-73557: vLLM is an inference and serving engine for large language models. From 0.20.2rc0 until 0.26.0, safe_load_prompt_embeds

vLLM (an AI inference and serving engine for large language models) versions 0.20.2rc0 to 0.26.0 have a vulnerability where concurrent requests to the chat API can bypass safety checks on prompt embeddings (pre-computed numerical representations of text input). This happens because the safety checking function uses a process-global state that can be exploited when multiple requests run simultaneously, potentially allowing invalid data to pass through even when safety features are enabled.