Frequently Asked Questions

Every request is evaluated across four detection modules: (1) PII Leakage — names, SSNs, credit cards, phone numbers; (2) Prompt Injection — jailbreaks, instruction overrides, role-play attacks; (3) Financial Hallucination — unauthorized refunds, pricing commitments; and (4) Brand Safety — offensive, biased, or discriminatory language.

We store only a SHA-256 hash of your API key — never the plaintext. Each request hashes the Bearer token and compares it against our database. This means if our database is compromised, your key cannot be reverse-engineered.

The Standard Engine (Sandox tier) uses Gemini 2.5 Flash — optimized for speed and cost. The Advanced Engine (Startup/Scale) uses Gemini 2.5 Pro — higher accuracy for nuanced, sophisticated threats. For production deployments serving real users, we recommend the Advanced engine.

This happens when the database trigger that creates your profile hasn't been applied yet. Have your administrator run the auth_trigger_fix.sql script in Supabase. This creates a profile row or backfills your account with $5.00 automatically.

You'll receive a 429 Too Many Requests response with a human-readable message. Requests do not count against your balance when rate-limited. Simply wait until the 60-second sliding window resets. Upgrade to Startup or Scale to remove all rate limits.

No. The proposed_text you send is processed in-memory and never persisted to disk. We log only metadata (timestamp, agent_id, risk_score, endpoint) for your billing and audit trail. This is by design for GDPR and SOC 2 compliance.

Guardrail.ai is completely LLM-agnostic. We evaluate text — we don't care what model produced it. You can use us with OpenAI, Anthropic, Gemini, Mistral, Llama, or any custom local model. We sit between the LLM output and your delivery layer.

The Custom Policy Engine lets you define compliance rules in plain English. For example: 'Never mention competitor products by name', 'Always include a disclaimer on medical advice', or 'Reject any response discussing political topics.' These rules are evaluated in addition to the default detection modules on every single request.

Our recommendation is a fail-open strategy for non-critical paths. If the guardrail API fails (server error, timeout), allow the response through with an internal alert — it's better to serve a potentially risky response than to break your user experience. For high-stakes flows (medical, financial), implement a fail-closed strategy and serve a safe fallback message instead.

Send a support ticket from your dashboard's Support tab with the full API response JSON and the original text. Our team reviews false positives and uses them to improve detection accuracy. We typically respond within 24 hours on Scale tier.