AI Gateway Guardrails
AI Gateway guardrails run inline on LLM requests and responses to detect and block prompt injection, sensitive data, harmful content, and off-topic.
| Where Can I Use This? | What Do I Need? |
|---|
- Prisma AIRS AI Gateway (Americas region)
|
- AI Gateway activated
- At least one LLM integration configured
|
Without guardrails, LLM traffic passes to and from your providers without inspection.
Guardrails solve this by running checks on every request before it reaches the model and
on every response before it reaches the caller. When a guardrail detects a violation, it
can deny the request, redact the offending content, or log the event and allow the traffic
through. Guardrails are referenced in routing configurations using
before_request_hooks (for request-side checks) and
after_request_hooks (for response-side checks), and are identified by a
pg- prefixed ID.
AI Gateway provides three tiers of guardrails:
Basic guardrails
Basic guardrails are built in and require no additional configuration. They perform
pattern-matching checks for common issues such as well-known prompt injection phrases,
explicit content, and simple keyword violations. Basic guardrails operate at low latency
and are suitable for high-throughput workloads where advanced detection is not required.
PRO guardrails
PRO guardrails are built-in advanced checks that use AI models to detect more
sophisticated issues. PRO guardrails include:
- No PII — Detects and optionally redacts personally identifiable information
(PII) in requests and responses, including names, addresses, phone numbers, and financial
identifiers.
- Content moderation — Identifies harmful, toxic, or policy-violating content
using a classification model.
- Topic control — Blocks requests or responses that discuss topics outside an
approved set, useful for constraining chatbots to a defined scope.
PRO guardrails add some processing latency. Each guardrail check consumes flex credits
in addition to the LLM call itself.
Partner guardrails
Partner guardrails integrate with third-party or Palo Alto Networks inspection services.
The
Prisma AIRS guardrail sends LLM requests and responses to the Prisma AIRS AI
Runtime Security API Intercept service for deep inspection, including detection of prompt
injection, data exfiltration attempts, and custom threat signatures. This requires a
separate AI Runtime Security deployment profile and is configured in Admin Settings →
Plugins. For configuration steps, see
Configure the Prisma AIRS Guardrail.
