Portkey Guardrails
Govern AI output with real-time input/output filtering – PII detection, content moderation, and custom logic enforced at the gateway layer

PII Detection & Redaction

Automatically detect and block names, email addresses, credit card numbers, and other personally identifiable information in prompts and responses before they reach external providers or end users.

• 20+ PII entity types supported.
• Scanning on both prompt input and model response output.
• Auto-redaction or hard block on detection.
• Full event log per triggered rule.

Content Moderation

Apply topic restrictions, toxicity filters, and hate-speech detection at the gateway level – no model fine-tuning required. Rules apply consistently across every model and every team using the gateway.

• Topic restriction rules per virtual key or team.
• Toxicity and hate-speech detection filters.
• Configurable sensitivity thresholds per rule.
• Works across all 1,600+ supported models.

Regex & Custom Rules

Define custom regex patterns or keyword blocklists to match proprietary data formats, internal terminology, or domain-specific risk patterns that general-purpose filters may not cover.

• Custom regex pattern matching.
• Keyword blocklists per team or environment.
• Domain-specific rule sets.
• Per-team and per-virtual-key configuration.

Input & Output Checks

Run guardrails on both the prompt going in and the model response coming out – providing end-to-end protection on every interaction rather than only one direction.

• Pre-request input scanning before the call reaches the provider.
• Post-response output scanning before the result reaches the user.
• Block or redact on rule trigger.
• Full audit trail per interaction.

Real-Time Alerts & Logging

Receive instant alerts when any guardrail fires. Every blocked event is logged with full context – request, rule, entity type, and timestamp – for compliance reporting and security incident review.

• Instant trigger alerts via dashboard, email, or webhook.
• Full blocked-event log with complete request context.
• Exportable logs for compliance and audit requirements.
• Per-rule analytics and hit-rate reporting in the dashboard.

Third-Party Safety Integrations

Connect Portkey Guardrails to external safety providers via a unified interface. Mix built-in guardrails with third-party checks in the same policy, applied consistently at the gateway layer regardless of which application or model is in use.

• Aporia, Pillar, and Patronus AI natively supported.
• Unified guardrail interface for built-in and partner checks.
• Pluggable architecture for additional safety vendors.
• Single policy configuration for all safety sources.

Protection at the Infrastructure Layer

Bolting safety onto individual applications is fragile and inconsistent. Because Guardrails sits inside the AI Gateway, policies are enforced centrally on every request – regardless of which team, framework, or model is making the call – without any changes to application code.

• Consistent enforcement across all teams and all models.
• No per-application safety code required.
• PII protection before data leaves your environment.
• Content moderation without model fine-tuning or redeployment.

Per-Team and Per-Key Configuration

Assign different guardrail policies to individual virtual keys, teams, or environments. A customer-facing application can enforce strict content rules while an internal research workload runs a lighter policy – all managed from one central dashboard.

Compliance and Audit Support

Every guardrail event is captured in a structured log with full request context. Logs are exportable for GDPR compliance reviews, SOC 2 audits, and internal security investigations. Portkey Guardrails operates within Portkey's SOC 2 Type II certified infrastructure, with a zero data retention option available.

Portkey Guardrails Specifications: