🚀 Veridex Protocol is live on testnet. Star us on GitHub →
SDKs
Agent Security

Agent Security (@veridex/agent-security)

A framework-agnostic security gateway that protects any AI agent — whether built with Veridex, LangChain, CrewAI, AutoGPT, or custom code — with pluggable security packs.

This package works standalone. It does not require @veridex/agents or any other Veridex package.

Installation

npm install @veridex/agent-security

Quick Start

In-Process Gateway

import { SecurityGateway, createDefaultPacks } from '@veridex/agent-security';
 
const gateway = new SecurityGateway({
  packs: createDefaultPacks(),
  defaultAction: 'block', // block | warn | allow
});
 
// Evaluate an agent action before execution
const result = await gateway.evaluate({
  type: 'tool_call',
  toolName: 'execute_sql',
  arguments: { query: 'DROP TABLE users; --' },
  agentId: 'data-agent',
  turnIndex: 3,
});
 
if (result.verdict === 'block') {
  console.log('Blocked:', result.reasons);
  // → ['Injection detected: SQL injection pattern in tool arguments']
}

Remote Gateway (HTTP)

Deploy as a standalone service:

// server.ts
import { createSecurityServer } from '@veridex/agent-security/server';
 
const server = createSecurityServer({
  packs: createDefaultPacks(),
  port: 4600,
  authToken: process.env.SECURITY_GATEWAY_TOKEN,
});
 
await server.start();

Connect from any agent:

import { SecurityClient } from '@veridex/agent-security';
 
const client = new SecurityClient({
  baseUrl: 'https://security.example.com',
  authToken: process.env.SECURITY_GATEWAY_TOKEN,
});
 
const result = await client.evaluate({
  type: 'tool_call',
  toolName: 'send_email',
  arguments: { to: 'user@example.com', body: emailContent },
  agentId: 'comms-agent',
});

Built-in Security Packs

The @veridex/agent-security package exports 10 pack factories. createDefaultPacks() returns 9 of them — endpointAllowlistPack is excluded from the default set because it requires an explicit allowlist and must be constructed by the caller.

PackIn createDefaultPacks()Detects
injectionDetectionPack✓Prompt injection, SQL injection, command injection in tool arguments
toolPoisoningPack✓Malicious tool chains, circular tool references, suspicious tool sequences
secretDetectionPack✓Leaked API keys, passwords, JWTs, private keys in arguments or outputs
budgetCeilingPack✓Spending exceeding configured limits per run or per day
shellCommandSafetyPack✓Dangerous shell commands (rm -rf, chmod 777, privilege escalation)
handoffSafetyPack✓Unsafe agent-to-agent handoffs (cycles, untrusted targets)
financialSafetyPack✓Financial transactions exceeding limits or targeting suspicious addresses
crossTurnAnomalyPack✓Behavioral anomalies across turns (goal drift, sudden capability escalation)
llmResponseGuardPack✓Filtering model responses for harmful or policy-violating content
endpointAllowlistPackHTTP requests to non-allowlisted domains (opt-in — requires allowedEndpoints)

Note on data sovereignty. A jurisdictional/PII enforcement pack is on the roadmap but is not yet shipped in @veridex/agent-security. Track progress in the roadmap.

Framework Adapters

LangChain

import { SecurityGateway } from '@veridex/agent-security';
import { VeridexSecurityCallback } from '@veridex/agent-security/adapters/langchain';
 
const gateway = new SecurityGateway({ packs: createDefaultPacks() });
const callback = new VeridexSecurityCallback(gateway);
 
// Attach to any LangChain agent
const agent = new AgentExecutor({
  agent: myAgent,
  tools: myTools,
  callbacks: [callback],
});

CrewAI

import { VeridexCrewAIGuard } from '@veridex/agent-security/adapters/crewai';
 
const guard = new VeridexCrewAIGuard(gateway);
 
// Wrap crew execution with security evaluation
const result = await guard.evaluate(crewAction);

Telemetry

import { TelemetryReporter } from '@veridex/agent-security';
 
const reporter = new TelemetryReporter({
  endpoint: 'https://telemetry.example.com/events',
  batchSize: 50,
  flushIntervalMs: 10_000,
});
 
const gateway = new SecurityGateway({
  packs: createDefaultPacks(),
  telemetry: reporter,
});
 
// Every evaluation is automatically recorded for observability

Key Types

interface SecurityVerdict {
  verdict: 'allow' | 'block' | 'escalate' | 'flag';
  reasons: string[];
  packId: string;
  confidence: number;
}
 
interface SecurityEvalResult {
  overall: 'allow' | 'block' | 'escalate';
  verdicts: SecurityVerdict[];
  metadata: Record<string, unknown>;
}
 
interface JurisdictionConfig {
  defaultJurisdiction: string;
  piiCategories: string[];
  jurisdictionRules: JurisdictionRule[];
  toolJurisdictions: Record<string, string>;
}

Related

Agent SDKAgents Framework