Mistral AI has released Shieldstral, a 3 billion parameter open-weights multimodal safety classifier designed to modernize content moderation through policy-adaptive inference. Available under the Apache 2.0 license, the model eliminates the traditional reliance on fixed harm taxonomies by framing safety evaluation as a dynamic question-answering task. Rather than requiring retraining for new deployment contexts, Shieldstral accepts plain-language safety policies at inference time, returning calibrated continuous safety scores from a single forward pass. The architecture operates on a straightforward yet highly effective principle. Each moderation request combines a document, a system instruction, and a user query expressed as a natural-language policy. The model processes this input and outputs softmax-normalized yes or no probabilities, enabling developers to set custom confidence thresholds or rank results dynamically. This approach unifies prompt classification, response moderation, refusal detection, and toxicity identification into a single interface that handles text, images, and combined modalities without architectural changes. In independent evaluations, Shieldstral matches or exceeds the performance of guardrail models up to seven times its size across text safety, refusal detection, multimodal benchmarks, and policy adaptability. Despite its compact scale, the model is engineered for production efficiency, requiring only a single 16GB NVIDIA GPU to operate. Its lightweight footprint and deterministic output format make it particularly suitable for latency-sensitive applications and resource-constrained deployments. The development of Shieldstral addressed several persistent challenges in safety model training. Public moderation datasets traditionally employ inconsistent taxonomies, annotation conventions, and labeling formats. Mistral engineers resolved this by converting heterogeneous sources into a unified instruction-query-document structure, adjusting per-dataset processing to preserve nuanced decision boundaries. To prevent memorization of predefined categories, the team employed contrastive training techniques, generating deliberately similar safe and unsafe pairs that force the model to learn precise policy distinctions rather than surface-level patterns. Visual safety data, inherently more difficult to synthesize, was augmented through mutation techniques, vision-language reranking, and high-quality negative sampling to minimize hallucinations and mislabeling. The final model emerged from a LoRA fine-tuning process followed by parameter merging of multiple specialized checkpoints, preserving both general instruction-following capabilities and fine-grained safety discrimination. Built using Mistral internal infrastructure, Shieldstral is positioned as a foundational component for adaptive moderation pipelines. Mistral cited its inaugural membership in the Open Secure AI Alliance, launched alongside NVIDIA and other industry partners, as a driving force behind the project. The company plans to expand the model capabilities in multilingual support, long-document robustness, and extended multimodal safety evaluations. By decentralizing safety taxonomies and placing policy definitions directly in the prompt, Shieldstral aims to provide a flexible, scalable alternative to static guardrail architectures, enabling developers to align model outputs with evolving regulatory and organizational standards without ongoing retraining overhead.