Safe by training and design
A useful model has to be trustworthy. We build safety into the training data, test it before every release, and check every request in production.
Train
Thousands of examples teach the model to refuse harmful requests, resist jailbreaks and ignore instructions hidden inside documents, without refusing questions that only sound sensitive.
Test
Before release, the model faces a red-team benchmark of harmful and jailbreak prompts. It ships only if it's at least as safe as the model it was built on.
Screen
In the API and chat, a separate safety model checks every prompt and every answer. Harmful requests get a polite refusal.
Govern
Clear usage rules, private customer data, and a way to report problems that we answer.
Honest when it doesn't know
Made-up answers are a safety problem too. When SORAEN answers from your documents, it cites the passage behind every fact, and it says plainly when the documents don't contain the answer. We measure this on every release.
Private by design
Documents you upload are visible only to your account and deleted when you delete them. Images and recordings are processed to answer you and not stored. We don't train on customer data. See the privacy policy.
What we can't promise
No AI model is perfectly safe or always right. Safety training and moderation reduce harmful outputs but can't rule them out, and answers can contain mistakes. Check important answers against their sources, and don't use SORAEN as the only safeguard for medical, legal, financial or safety-critical decisions.
Found a safety issue?
Tell us about harmful outputs, jailbreaks or security problems. We read every report.
founder.soraen.ai@gmail.com ยท Usage rules are in our terms.