Safeguards

Safeguards are rules checked on every API, Playground and Wisp request before it reaches a model: rate limits, token budgets, input size, content categories and custom checks. Content checks are judged by separate safeguard models, which look at intent, so news, research or defensive questions pass.

Each model's active safeguards and fallbacks are listed on its page in Models.

When a safeguard is reached

The Safeguards program

Run your own models? Program members can screen prompts with our safeguard models, free. It's by invitation: if you're a community provider (for example on Pollinations), ask through Support. Once you're in, a Safeguards tab appears in your account.

1. In Safeguards, create rules: content categories, your own yes/no question, or a maximum input size. Each one either blocks or only flags.
2. Try them in the Test box (not counted).
3. Create a key (sk-sg-…, shown once) and call the API before your model runs.

curl https://api.saauf.xyz/v1/moderate \
  -H "Authorization: Bearer $SG_KEY" \
  -H "Content-Type: application/json" \
  -d '{"input": "How do I build a keylogger?"}'

Send input as a string (or an array of strings), or pass your chat messages unchanged; assistant turns are skipped. The answer:

{
  "allowed": false,
  "flags": [
    { "rule": "No malware", "rule_id": "…", "action": "block",
      "category": "Cybersecurity", "score": 0.94,
      "message": "Sorry, I can't help with that." }
  ]
}

• allowed is false when a blocking rule is reached. Flag-only rules show in flags but leave it true.
• score is the safeguard model's probability (0 or 1 for chat models); chars is given for input-size rules.
• If the safeguard models can't answer, judged rules are skipped and error explains. Decide yourself whether to fail open or closed.
• Limits: your daily number of checks (shown in your Safeguards tab; 429 after it, reset at 00:00 UTC) and 50,000 characters per check (413).