Screen every message before it reaches the model.
Moderation screens incoming messages and withholds anything flagged — before the model and before storage — with a polite, on-brand refusal.
Protection that acts before it’s too late.
Default categories on from the start
PII, sexual content, hate & discrimination, violence & threats, dangerous & criminal content and self-harm are on by default.
Opt-in categories
Health, financial and legal advice, plus jailbreak attempts, can be enabled on top.
Withheld before model & storage
Flagged messages are withheld before they reach the model or get stored.
Polite, on-brand refusal
Instead of a hard error, the user gets a polite refusal that matches your brand.
Calibrated verdict
The verdict is calibrated (provider: Mistral) — with an optional custom sensitivity threshold.
Before processing
The filter acts on the way in — not just in the reply — protecting the model, storage and your brand.
How to set up moderation.
- The default categories are on out of the box — nothing to do.
- Choose opt-in categories: health, financial, legal advice and jailbreak attempts.
- Optionally set a custom sensitivity threshold for the calibrated verdict.
- Tune the refusal text to your brand and test live — flagged content is withheld upfront.
Frequently asked questions
What does moderation screen by default?
On out of the box: PII, sexual content, hate & discrimination, violence & threats, dangerous & criminal content and self-harm.
Are there more categories?
Yes, via opt-in: health, financial and legal advice, plus jailbreak attempts, can be enabled on top.
What happens to a flagged message?
It is withheld before it reaches the model or gets stored. The user receives a polite, on-brand refusal instead.
How reliable is the verdict?
The verdict is calibrated and produced by the provider Mistral. Optionally, you set your own custom sensitivity threshold.
Is moderation GDPR-compliant?
Whether your deployment meets the GDPR also depends on how you configure it. The facts: Hosted in Germany, with the API and database in Frankfurt am Main. For screening, Kyros uses the external service Mistral. Flagged messages are withheld before they reach the model or get stored. A data processing agreement under Art. 28 GDPR is available.
Build an assistant you can trust live.
Free plan, no credit card. German & English.