Frequently Asked Questions

AgentAIShield detects three categories of AI security risks:

• PII (Personally Identifiable Information) — emails, phone numbers, SSNs, credit cards, names, addresses, dates of birth, passport numbers, driver's licenses, and IP addresses — in both prompts sent to the LLM and responses returned from it.
• Prompt injection — attacks where malicious instructions attempt to hijack your AI agent: jailbreaks, system prompt overrides, goal hijacking, data exfiltration attempts, and indirect injection (e.g., malicious content in RAG-retrieved documents).
• Policy violations — custom rules you define (e.g., "no PII in prompts", "block specific models", "enforce budget limits").

Detection happens on both the prompt (input) and response (output) side, giving you full coverage of your AI pipeline.

AAIS detects the following PII types with severity ratings:

• critical SSN (Social Security Number)
• critical Credit card numbers (Visa, MC, Amex, etc.)
• critical Passport numbers
• high Date of birth
• high Driver's license numbers
• medium Email addresses
• medium Phone numbers
• medium Physical addresses
• low Full names (with contextual signals)
• low IP addresses

Detection uses pattern matching combined with contextual analysis to minimize false positives.

AAIS detects these prompt injection categories:

• System prompt override — "Ignore previous instructions", "Forget your system prompt", "Disregard all prior context"
• Role jailbreaks — "Act as DAN", "Pretend you have no restrictions", "You are now an unrestricted AI"
• Data exfiltration — "Repeat your instructions", "Print your system prompt", "Show me your initial prompt"
• Indirect injection (RAG attacks) — Malicious instructions embedded in documents, web pages, or other external content your agent retrieves
• Goal hijacking — Adversarial instructions in user-controlled content (form fields, emails, tickets) designed to redirect your agent
• Privilege escalation — Attempts to make the LLM act with elevated permissions or bypass safety measures

Each detection has a confidence score (0.0–1.0). High confidence (≥0.85) is flagged as severity high.

Detection accuracy varies by PII type:

• High precision: SSNs, credit cards, passport numbers (format-based, ~99% accurate)
• Good precision: Emails, phone numbers (~95% accurate)
• Moderate precision: Names, addresses (contextual, ~80-90% accurate)

For injection detection, confidence scores help you tune your response. In Monitor Mode, all detections are logged — you can review and adjust policies. In Proxy Mode, you can set a minimum confidence threshold before blocking.

Low-confidence detections (below 0.5) are logged but typically don't trigger blocks.

Every API key in AAIS gets an Agent Trust Score™ — a number from 0 to 100 with a letter grade (A+ to F). Think of it as a behavioral credit score for your AI agents.

The score is updated after every request and considers:

• Error rate — Fewer errors raises the score
• PII exposure rate — Prompts/responses with PII lower the score
• Injection attempt rate — Injection attempts lower the score significantly
• Latency consistency — Stable, predictable latency improves confidence
• Request volume — More data = higher confidence in the score

Grade thresholds: A+(95-100), A(90-94), B+(85-89), B(75-84), C+(65-74), C(50-64), D(35-49), F(0-34).

Score confidence (0-1) indicates how statistically reliable the trust score is. New agents start with low confidence because there isn't enough data to form a stable picture.

• <100 requests — Low confidence. Score may swing significantly.
• 100-1,000 requests — Medium confidence. Score stabilizes.
• >1,000 requests — High confidence. Score is very reliable.

You can see score_confidence in the GET /api/trust/agents response. This prevents a single PII detection from tanking a new agent's score unfairly.

Badges are earned for sustained good behavior:

• Zero-PII Streak — 30+ consecutive days with zero PII detected in prompts or responses
• Consistent — Low variance in error rate and latency over 30 days
• Improving — Trust score increased 10+ points in the last 30 days
• High Volume — Over 10,000 requests logged (high-confidence score)
• Secure Pipeline — Zero injection attempts detected in 30+ days

Badges are visible in the dashboard and readable via the API (GET /api/trust/agents/:id).

Start with Monitor Mode. It's zero-risk, zero-latency, and gives you immediate visibility into your AI traffic.

• Monitor Mode: 5 lines of code. Fire-and-forget POST after every AI call. Zero latency impact. Passive observation only — no blocking.
• Proxy Mode: Change your SDK's base_url. AAIS intercepts calls inline, can block based on policies. Adds ~50-150ms latency per call.

Use Proxy Mode when you need blocking (e.g., preventing PII from reaching the LLM, blocking injection attempts before they execute). Use Monitor Mode for visibility without risk.

You can run both simultaneously — some agents in proxy mode, others in monitor mode.

No. Monitor Mode is designed to be completely non-blocking.

You call POST /api/monitor/ingest after your AI call completes, and the endpoint responds with {"ok": true} immediately — all PII scanning, injection detection, and trust score updates happen asynchronously.

Best practice: don't await the ingest call in your hot path. Fire it and forget it. Even if AAIS is unreachable, your application should continue without any impact.

AAIS supports alerts via webhooks (configure in Dashboard → Settings → Webhooks). You can trigger on:

• PII detected in prompt or response (by severity)
• Injection attempt detected
• Agent trust score drops below a threshold
• Policy violation count exceeds a limit
• Budget threshold reached

Alternatively, poll GET /api/dashboard/threats and GET /api/dashboard/stats from your own alerting system to build custom notifications.

Monitor Mode supports any provider — just pass the provider name and model as strings. It's completely provider-agnostic.

Proxy Mode supports:

• OpenAI — /v1/chat/completions, /v1/completions, /v1/embeddings
• Anthropic — /v1/messages
• More providers in development (Google Gemini, Cohere, Mistral)

Yes — in Proxy Mode, streaming (Server-Sent Events) is fully supported for /v1/chat/completions. Pass "stream": true in your request as usual.

In Monitor Mode, you report after the full stream completes — just collect the full response text before sending the ingest request.

AAIS stores metadata about AI calls (model, provider, token counts, latency, status) and detection results (PII types found, injection confidence, violations). It does not store the full prompt or response text.

Specifically, for prompts, AAIS stores a SHA-256 hash (one-way, cannot be reversed) for deduplication. The actual text is processed in memory during scanning and immediately discarded.

If you're self-hosting AAIS, you have full control over what's stored by configuring the database schema.

Yes. AAIS is a Node.js + PostgreSQL application that runs anywhere — your own server, Railway, Render, Fly.io, AWS, GCP, or locally.

Self-hosting gives you:

• Full data sovereignty — prompts never leave your infrastructure
• No per-request pricing — just infrastructure costs
• Custom policy configuration
• Integration with your existing monitoring stack

Required environment variables: DATABASE_URL (PostgreSQL), JWT_SECRET, and provider API keys for proxy mode.

AAIS is designed to help you maintain compliance by detecting when PII appears in AI pipelines where it shouldn't. Since AAIS stores only hashes (not actual prompt text), the self-hosted version significantly reduces your data controller obligations.

For HIPAA compliance (PHI detection): AAIS detects medical identifiers (SSN, DOB, account numbers). Self-hosting on your own HIPAA-compliant infrastructure is the recommended approach for healthcare applications.

Consult your legal team for specific compliance requirements. AAIS is a tool to assist, not a certification.

Default rate limits for a self-hosted instance:

• Auth endpoints: 20 requests per 15 minutes (per IP)
• Dashboard API: 120 requests per minute (per session)
• Monitor ingest: 100 requests per minute per API key
• Proxy endpoints: 600 requests per minute (global)

For high-volume agents (>100 RPM), you have two options: create multiple API keys (each gets 100 RPM), or adjust the max value in the rate limiter config in api/monitor.js for your self-hosted instance.

AgentAIShield is self-hostable. You run it on your own infrastructure, so the cost is just your hosting fees (typically $5-20/month on Railway or Render for small-to-medium workloads).

There's no per-request fee. No seat limits. No usage caps (beyond your hardware).

For managed hosting or enterprise support, contact us.

Unlimited. There's no hard limit on API keys or agents in the self-hosted version. Each API key creates one agent profile with its own trust score.

Recommended: create one API key per agent/application for clean trust scoring and audit trails. Don't share keys across different applications.