Directive provides security guardrails and compliance tooling for AI agent systems. Detect threats at the input layer, audit every operation, and handle data subject requests.

Security Features

Feature	Page	Threat Addressed
PII Detection	Input/output scanning	Personally identifiable information leaking to/from agents
Prompt Injection	Input validation	Jailbreaks, instruction overrides, encoding evasion
Audit Trail	Observability	Tamper-evident logging of every system operation
GDPR/CCPA	Data governance	Right to erasure, data export, consent tracking, retention

Defense in Depth

Apply multiple layers of protection:

User Input
  → Prompt Injection Detection  (block attacks before they reach agents)
  → PII Detection               (redact sensitive data from input)
  → Agent Execution              (safe to process after filtering)
  → Output PII Scan             (catch any data leaks in responses)
  → Audit Trail                 (log every operation for compliance)

Quick Setup

import { createAgentOrchestrator } from '@directive-run/ai';
import { createEnhancedPIIGuardrail, createPromptInjectionGuardrail } from '@directive-run/ai';

const orchestrator = createAgentOrchestrator({
  runner: myRunner,

  // Input guardrails run in order before each agent invocation
  guardrails: {
    input: [
      // First line of defense: block injection attacks
      createPromptInjectionGuardrail({ strictMode: true }),

      // Second pass: redact any PII that slipped through
      createEnhancedPIIGuardrail({ redact: true }),
    ],
  },
});

When to Use What

Scenario	Feature
User-facing chatbot	PII detection + prompt injection + audit trail
Internal tool	Audit trail + GDPR compliance
Healthcare/finance	All four features
Development/testing	Audit trail only

Server vs. Browser

All security features use globalThis.crypto.subtle (Web Crypto API) and work in both environments:

Feature	Server (Node 18+)	Browser	Primary Use
PII Detection	Yes	Yes	Both – input/output scanning
Prompt Injection	Yes	Yes	Both – input validation
Audit Trail	Yes	Yes	Primarily server – tamper-evident logs, SIEM export
GDPR/CCPA Compliance	Yes	Yes	Primarily server – data export, deletion certificates

Node 18+ is required for crypto.subtle. No Node-specific imports are used – the same code runs in Deno and Bun as well.

Audit trails and compliance tooling are most useful on the server where you control the data layer. PII detection and prompt injection prevention are equally valuable in both environments. See the Server (Node.js) example for a full API demonstrating all four features together.

Next Steps

Start with safety – PII Detection is the most common first step
Add attack prevention – Prompt Injection for user-facing apps
Compliance requirements? – GDPR/CCPA for regulated industries

Security & Compliance Overview