@superagent-ai/mastra

Superagent security processors for Mastra AI agents. Protect your AI applications with threat detection (Guard) and PII redaction (Redact).

Built on top of @superagent-ai/safety-agent SDK.

Installation

npm install @superagent-ai/mastra

Features

• Guard Processor - Detect and block prompt injection, system prompt extraction, and data exfiltration attempts
• Redact Processor - Automatically remove PII/PHI from user inputs before processing
• TypeScript - Full type safety with exported types
• Multi-Provider Support - Use various LLM providers (OpenAI, Anthropic, Google, etc.) through the safety-agent SDK

Environment Variables

The processors require specific environment variables depending on which features you use:

# Required for all processors
SUPERAGENT_API_KEY=your-superagent-api-key

# Required for Redact processor (depends on which model you choose)
ANTHROPIC_API_KEY=your-anthropic-api-key   # If using anthropic/* models
OPENAI_API_KEY=your-openai-api-key         # If using openai/* models
GOOGLE_API_KEY=your-google-api-key         # If using google/* models

Note: The Guard processor uses Superagent's hosted model by default and only requires SUPERAGENT_API_KEY. The Redact processor requires an additional API key for the LLM provider you choose.

Quick Start

import { Agent } from "@mastra/core/agent";
import {
  SuperagentGuardInputProcessor,
  SuperagentRedactInputProcessor,
} from "@superagent-ai/mastra";

const agent = new Agent({
  name: "secure-agent",
  instructions: "You are a helpful assistant.",
  model: "anthropic/claude-3-5-haiku-20241022",

  inputProcessors: [
    // Block malicious inputs (prompt injection, etc.)
    new SuperagentGuardInputProcessor({
      apiKey: process.env.SUPERAGENT_API_KEY!,
    }),
    // Strip PII from user inputs before processing
    new SuperagentRedactInputProcessor({
      apiKey: process.env.SUPERAGENT_API_KEY!,
      model: "anthropic/claude-3-5-haiku-20241022", // Requires ANTHROPIC_API_KEY
    }),
  ],
});

Processors

SuperagentGuardInputProcessor

Analyzes user inputs for security threats before they reach your agent. Uses Superagent's optimized guard model by default.

new SuperagentGuardInputProcessor({
  apiKey: "your-api-key",
  model: "superagent/guard-1.7b", // optional - uses Superagent's default guard model
  systemPrompt: "Custom instructions for classification", // optional
});

Detects:

• Prompt injection attempts
• System prompt extraction attacks
• Data exfiltration attempts

Response when blocked: When a threat is detected, the processor triggers a tripwire and the request is aborted with the violation types (e.g., prompt_injection, system_prompt_extraction).

SuperagentRedactInputProcessor

Removes sensitive information from user inputs before the agent processes them.

new SuperagentRedactInputProcessor({
  apiKey: "your-api-key",
  model: "anthropic/claude-3-5-haiku-20241022", // optional - defaults to anthropic/claude-3-5-haiku-20241022
  entities: ["email addresses", "social security numbers", "phone numbers"], // optional
  rewrite: false, // optional - when true, rewrites text contextually instead of using placeholders
});

Default entities redacted:

• SSNs, Driver's License, Passport Numbers
• API Keys, Secrets, Passwords
• Names, Addresses, Phone Numbers
• Emails, Credit Card Numbers

Example output:

Input:  "My email is john@example.com and SSN is 123-45-6789"
Output: "My email is <EMAIL_REDACTED> and SSN is <SSN_REDACTED>"

Configuration

Common Options

Option	Type	Required	Description
apiKey	string	Yes	Your Superagent API key
model	string	No	Model in "provider/model" format

Guard Options

Option	Type	Required	Default	Description
systemPrompt	string	No	-	Custom instructions to steer classification behavior

Redact Options

Option	Type	Required	Default	Description
model	string	No	anthropic/claude-3-5-haiku-20241022	Model for redaction
entities	string[]	No	Standard PII	Custom entity types to redact
rewrite	boolean	No	false	Rewrite text contextually instead of using placeholders

Supported Models

Use the provider/model format when specifying models:

Provider	Model Format	Required Env Variable
Superagent	superagent/{model}	None (default for guard)
Anthropic	anthropic/{model}	ANTHROPIC_API_KEY
OpenAI	openai/{model}	OPENAI_API_KEY
Google	google/{model}	GOOGLE_API_KEY
Groq	groq/{model}	GROQ_API_KEY
Fireworks	fireworks/{model}	FIREWORKS_API_KEY
AWS Bedrock	bedrock/{model}	AWS_BEDROCK_API_KEY
OpenRouter	openrouter/{provider}/{model}	OPENROUTER_API_KEY
Vercel AI Gateway	vercel/{provider}/{model}	AI_GATEWAY_API_KEY

Direct SDK Usage

For advanced use cases, you can use the safety-agent SDK directly:

import { createClient } from "@superagent-ai/mastra";

const client = createClient({ apiKey: process.env.SUPERAGENT_API_KEY });

// Guard - Classify input as safe or unsafe
const guardResult = await client.guard({
  input: "user message to analyze",
});

if (guardResult.classification === "block") {
  console.log("Blocked:", guardResult.violation_types);
}

// Redact - Sanitize sensitive content
const redactResult = await client.redact({
  input: "My email is john@example.com",
  model: "anthropic/claude-3-5-haiku-20241022",
});

console.log(redactResult.redacted);
// "My email is <EMAIL_REDACTED>"

API Reference

For more information about the underlying APIs:

• Safety Agent SDK Documentation
• @superagent-ai/safety-agent on npm

License

MIT