Mistral's new model reads your safety policy at runtime
Shieldstral is a 3B open-weight safety classifier that takes a plain-language moderation policy as input, no retraining needed to change the rules.
Mistral released Shieldstral, a 3-billion-parameter open-weight safety classifier that Mistral says outperforms models up to seven times its size. Rather than training a separate classifier for every moderation rule, Shieldstral takes a plain-language policy as a question at inference time and returns a calibrated safety score, covering both text and images through one interface. It runs on a single 16GB GPU and ships under Apache 2.0.
Mistral released the model as a founding contribution to the Open Secure AI Alliance, the Nvidia-led coalition of more than 50 companies building shared open-source defenses for AI systems.
What it means for operators
Shieldstral turns content-policy changes into a prompt-editing task instead of a retraining job, which matters if your product has to enforce different moderation rules across markets, customer tiers, or regulatory regimes and you’ve been maintaining multiple classifiers to do it. It also adds a concrete open-weight tool to the Open Secure AI Alliance’s stated goal of giving defenders inspectable, self-hosted security tooling instead of a closed vendor API you can’t audit, alongside n8n’s own recent work on scoped, per-agent MCP permissions.