Mistral has released a new AI safety tool called Shieldstral. The model checks whether content is safe or unsafe based on rules written in plain language. It was announced on August 4, 2026.
Shieldstral is small compared to other safety models. It has 3 billion parameters, yet Mistral says it matches or beats guardrail models up to seven times larger.
The company built Shieldstral to solve a common problem. Most safety classifiers are trained on a fixed list of harm categories. If a company wants to change its rules, it usually has to retrain the whole model.
How Shieldstral Works
Shieldstral takes a different approach. Users write a policy as a question, such as asking if content promotes violence or if an image is safe for a minor to see.
The model reads the question along with the content and returns a yes or no answer. It also gives a confidence score for that answer.
This means one version of the model can adapt to different rules for different products. A cybersecurity tool and a mental health app might have different safety needs, and Shieldstral can adjust without new training.
The system works on text, images, and combinations of both. It can check user prompts, AI responses, or full conversations between the two.
Mistral says this single design covers several jobs at once. That includes filtering harmful prompts, moderating responses, detecting when an AI improperly refuses a request, and flagging toxic language.
Training Data And Design
Mistral says the model's performance comes mostly from how the training data was built. Public safety datasets often use different labels and category systems, so the team converted them into one shared format.
The team also created pairs of similar examples where only one breaks a specific rule. This was meant to teach the model to tell close policies apart instead of memorizing a fixed list.
Images posed a separate challenge. Unlike text, harmful images cannot be generated by an AI language model for training purposes.
To work around this, Mistral used general image datasets as safe examples and checked labels using a separate vision based model. This was done to reduce mistakes in the training data.
The final model combines several trained versions into one. Mistral merged a version trained on public data with another trained on generated policy examples, plus the original base model.
Shieldstral was built using Forge, which is Mistral's internal platform for training and evaluating models. The company says Forge handled the technical infrastructure so the team could focus on the training data itself.
Mistral released Shieldstral under the Apache 2.0 license, meaning anyone can download and use the model. It is available now on Mistral's platform.
The release comes as Mistral joins the Open Secure AI Alliance, a group that includes NVIDIA and other companies focused on AI safety tools. Mistral said it plans to keep working on multilingual support and broader image and video safety coverage going forward.
The model requires only a single 16GB graphics card to run, which Mistral says makes it accessible for smaller companies. Full technical details were published in a report alongside the release.
