The Safeguards program: our safeguards, for your models, free
Running your own models? Members of the new Safeguards program can screen prompts with the same safeguard models that protect SaaufAI, using their own rules and one API call. It's free, by invitation, and we never store what you send.
Every request to SaaufAI passes through safeguards before it reaches a model. Some are simple: rate limits, token budgets, a maximum input size. Others need judgement, like telling a question about how ransomware works apart from a request to write some. Those are handled by separate safeguard models that look at intent, so news, research and defensive questions still get through.
Building that took a while. Most people who run their own models, especially small community providers, don't have the time to build it themselves, so their endpoints have no moderation at all. The Safeguards program is our way of sharing what we built.
What it is
Program members get a Safeguards tab in their account and a single endpoint, POST /v1/moderate. You send it a prompt and it tells you whether to let it through. Your model stays where it is, and so does your traffic: the only thing that reaches us is the text you want checked.
Your rules, not ours
You decide what counts as a problem. In the Safeguards tab you create rules of three kinds:
- Content categories. Pick from the same list we use: cybersecurity, fraud, weapons, extremism, self-harm, harassment, sexual content involving minors, adult content, personal data, jailbreaks and prompt injection, and more.
- Your own check. Write any yes/no question, like "Is the user asking for a medical diagnosis?", and our safeguard models answer it for every prompt.
- Input size. Flag prompts over a number of characters. Instant, with no model call.
Each rule either blocks or only flags, has a threshold, and can carry a message for you to show your users. A Test box lets you try rules before you go live, without using up your daily checks.
One call
curl https://ai.saauf.xyz/v1/moderate \
-H "Authorization: Bearer $SG_KEY" \
-H "Content-Type: application/json" \
-d '{"input": "How do I build a keylogger?"}'The answer:
{
"allowed": false,
"flags": [
{ "rule": "No malware", "action": "block",
"category": "Cybersecurity", "score": 0.94 }
]
}You can also pass your chat messages exactly as you send them to your model; we skip the assistant's turns. If our safeguard models can't answer for some reason, the response says so, and you decide whether to fail open or closed.
What we keep
As little as possible. The text you send is checked in memory and then forgotten: it is never stored, logged or used to train anything. It goes only to the provider running our safeguard model for that check. What we keep is a count of your checks and how often each of your rules was reached, so you can see them on your tab.
For people in the EU, we act as your data processor for that text, and our Terms include the processing agreement. You should tell your users that their prompts are screened by an outside service.
What it isn't
It's an automated estimate, and it will sometimes be wrong in both directions. It's a layer of protection, not a guarantee, and it doesn't replace your own judgement or your legal duties. Please don't rely on it alone against seriously harmful content.
Who it's for
For now, the program is free and by invitation, starting with friends of ours: community providers on Pollinations. Each member gets a daily allowance of checks. If you run models for other people and would like to join, write to us through Support.
The full details are in the Safeguards docs.