Guardrails & Compliance

Budget limits, PII redaction, prompt injection detection, and compliance profiles for enterprise-grade AI safety.

Overview

AllRoutes Guardrails provide a policy layer between your application and upstream AI providers. Guardrails enforce budget limits, filter harmful content, redact sensitive data, and ensure compliance with regulatory frameworks -- all without modifying your application code.

Budget Guardrails

Prevent runaway costs with per-key and per-organization spending controls:

Daily budget -- maximum spend per calendar day
Monthly budget -- maximum spend per calendar month
Per-request cost cap -- reject requests estimated to exceed a cost threshold

Budget limits are configured on API keys or at the organization level in the dashboard.

When a budget is exhausted, the API returns 402 Payment Required with a descriptive error message.

Model and Provider Allowlists

Restrict which models and providers your keys can access:

{
  "allowed_models": ["gpt-4o", "claude-sonnet-4-20250514"],
  "blocked_providers": ["xai"],
  "require_byok": true
}

allowed_models -- whitelist specific model IDs
blocked_providers -- blacklist entire providers
require_byok -- only allow requests through your own provider keys (Zero Data Retention)

Zero Data Retention (ZDR)

When ZDR is enforced, AllRoutes ensures:

No request or response data is logged or stored
Only BYOK provider keys are used (providers with ZDR agreements)
Cache is disabled for ZDR requests
Broadcast destinations receive metadata only (no prompt/response content)

Enable ZDR at the organization level or per-key.

PII Redaction

AllRoutes can automatically detect and redact Personally Identifiable Information in both requests and responses:

Bidirectional Redaction

Inbound: PII is detected in user messages and replaced with placeholders before sending to the provider
Outbound: Placeholders are restored with the original values in the response returned to your application

This means the provider never sees actual PII, while your application receives the complete response with real data.

Detected PII Types

Type	Example	Placeholder
Email addresses	john@example.com	`[EMAIL_1]`
Phone numbers	+1-555-0123	`[PHONE_1]`
Social Security Numbers	123-45-6789	`[SSN_1]`
Credit card numbers	4111-1111-1111-1111	`[CARD_1]`
IP addresses	192.168.1.1	`[IP_1]`
Names (detected via NER)	John Smith	`[NAME_1]`
Physical addresses	123 Main St, NYC	`[ADDRESS_1]`

Enable PII redaction via the plugins array or as an organization default:

{
  "model": "gpt-4o",
  "messages": [{"role": "user", "content": "My email is john@example.com"}],
  "plugins": [{"id": "pii-redactor"}]
}

Prompt Injection Detection

AllRoutes scans incoming messages for prompt injection attempts -- malicious instructions designed to override the system prompt or extract sensitive information.

When an injection is detected:

The request is blocked with a 400 response
An alert is sent to your configured webhook
The attempt is logged for security review

Enable via plugins:

{
  "plugins": [{"id": "injection-detector"}]
}

Detection covers common injection patterns: ignore-previous-instructions, role-play jailbreaks, encoding-based bypasses, and delimiter injection.

Toxicity Filtering

Filter harmful, offensive, or inappropriate content in both prompts and responses:

{
  "plugins": [{"id": "toxicity-filter"}]
}

The toxicity filter scores content across categories (hate speech, harassment, self-harm, sexual content, violence) and blocks requests or responses that exceed configurable thresholds.

Compliance Profiles

AllRoutes offers pre-built compliance profiles that bundle multiple guardrails into a single configuration:

Profile	Includes
HIPAA	ZDR, PII redaction, audit logging, US-region routing, encrypted transit
SOC 2	Audit logging, access controls, encryption, incident alerting
GDPR	EU-region routing, PII redaction, data minimization, right-to-erasure support
EU AI Act	Model transparency, risk classification, human oversight hooks
PCI-DSS	Credit card redaction, ZDR, encrypted transit, access logging
FERPA	Student data redaction, ZDR, audit logging, US-region routing

Apply a compliance profile per request:

{
  "model": "gpt-4o",
  "messages": [{"role": "user", "content": "..."}],
  "guardrail": "hipaa"
}

Or set a default compliance profile for your organization in the dashboard under Settings > Compliance.

Guardrail Response Headers

Header	Description
`X-AllRoutes-Guardrail`	Active guardrail profile
`X-AllRoutes-PII-Redacted`	Number of PII entities redacted
`X-AllRoutes-Injection-Score`	Injection detection confidence (0-1)
`X-AllRoutes-Toxicity-Score`	Toxicity score (0-1)

Best Practices

Start with budget guardrails -- always set daily/monthly limits on production keys
Enable PII redaction for any user-facing application handling personal data
Use compliance profiles to simplify regulatory adherence
Monitor guardrail metrics in the dashboard for false-positive tuning
Layer guardrails -- combine multiple protections for defense in depth