Mistral Launches Shieldstral: An Open Safety Classifier for AI Content Moderation

French AI lab Mistral introduced Shieldstral on Tuesday, a three-billion-parameter multimodal safety classifier designed to streamline AI safety guardrails. Unlike traditional models that require retraining to update safety policies, Shieldstral accepts natural language instructions at inference time, allowing users to define custom safety policies without model retraining. The system produces a continuous safety score rather than binary checks, giving deployers more granular control over risk thresholds. Mistral claims the model matches the performance of OpenAI's GPT-OSS-Safeguard, a twenty-billion-parameter model, on text safety benchmarks while running on as little as 16GB of GPU memory, significantly reducing computational costs. Released as open-weight under the Apache 2.0 license, Shieldstral reflects Mistral's strategy of building practical tools across the AI stack rather than focusing solely on frontier model development. The move aligns with the company's recent joining of the Open Secure AI Alliance, a newly launched coalition focused on open-source safety tools.