A new AI safety model developed by Mistral has demonstrated comparable performance to larger models in certain benchmarks, while requiring significantly less computational resources. This model, called Shieldstral, uses natural language questions to evaluate AI inputs and outputs for safety concerns. One of its key features is the ability for operators to define their own safety criteria at runtime, rather than relying on pre-defined categories. Additionally, Shieldstral can be run locally, without the need for a third-party service. This development has implications for the deployment of AI systems in various industries, where safety and efficiency are crucial considerations.