Mistral AI Releases Shieldstral: A Policy-Adaptive Multimodal Safety Classifier

Shieldstral is a 3B open-weights multimodal safety classifier released by Mistral AI under the Apache 2.0 license. Unlike traditional guardrail models that bake a fixed taxonomy of harm categories into their weights, Shieldstral takes a novel approach: users write the safety policy as a plain-language question at inference time, and the model returns a calibrated safety score. This formulation unifies prompt classification, response moderation, refusal detection, and toxicity detection into a single problem, and it works for both text and images without retraining. The model matches or outperforms open guard models up to 7x its size across text safety, refusal detection, policy adaptability, and multimodal benchmarks. It runs efficiently on a single 16GB NVIDIA GPU. The training process involved unifying heterogeneous public safety datasets, teaching discrimination through contrastive pairs generated by an LLM, grounding safety in images with vision-language reranking, and combining complementary checkpoints via LoRA and SLERP. Mistral AI built Shieldstral end-to-end on its Forge platform. The company is continuing to work on multilingual coverage, longer-document robustness, and broader multimodal safety. Mistral AI is also an inaugural member of the Open Secure AI Alliance with NVIDIA and other organizations.