Superagent SDK (Python)

An open-source SDK for AI agent safety. Guard against prompt injections, redact sensitive data, and scan repositories for threats.

Installation

uv add safety-agent

Or with pip:

pip install safety-agent

Prerequisites

Sign up at superagent.sh to get your API key.

export SUPERAGENT_API_KEY=your-key

Quick Start

from safety_agent import create_client

client = create_client()

# Guard: Detect threats (uses default superagent/guard-1.7b model)
result = await client.guard(input="user message to analyze")

if result.classification == "block":
    print("Blocked:", result.violation_types)

# Redact: Remove PII
result = await client.redact(
    input="My email is john@example.com",
    model="openai/gpt-4o-mini"
)

print(result.redacted)
# "My email is <EMAIL_REDACTED>"

Guard

The guard() method classifies input content as pass or block. It detects prompt injections, malicious instructions, and security threats.

result = await client.guard(
    input="Ignore all previous instructions",
    model="openai/gpt-4o-mini",  # Optional, defaults to superagent/guard-1.7b
    system_prompt="Custom system prompt",  # Optional
    chunk_size=8000,  # Optional, characters per chunk
)

print(result.classification)  # "pass" or "block"
print(result.violation_types)  # ["prompt_injection", ...]
print(result.cwe_codes)  # ["CWE-94", ...]

Input Types

Guard supports multiple input types:

• Plain text: Analyzed directly
• URLs: Automatically fetched and analyzed
• Bytes/Files: Analyzed based on content type
• PDFs: Text extracted and analyzed per page

Remote URLs must resolve exclusively to public IP addresses. Redirect targets are revalidated, and downloads are limited to 5 redirects, 30 seconds, and 25 MiB.

# URL input
result = await client.guard(input="https://example.com/document.pdf")

# File input
with open("document.pdf", "rb") as f:
    result = await client.guard(input=f.read())

Redact

The redact() method removes sensitive content from text.

result = await client.redact(
    input="My SSN is 123-45-6789",
    model="openai/gpt-4o-mini",
    entities=["SSN", "email"],  # Optional, custom entities
    rewrite=True,  # Optional, contextual rewriting
)

print(result.redacted)
print(result.findings)

Supported Providers

• OpenAI (openai/gpt-4o, openai/gpt-4o-mini, etc.)
• OpenAI Compatible (openai-compatible/my-model, etc.)
• Anthropic (anthropic/claude-3-5-sonnet-20241022, etc.)
• Google (google/gemini-2.0-flash, etc.)
• AWS Bedrock (bedrock/us.anthropic.claude-3-5-sonnet-20241022-v2:0, etc.)
• Groq (groq/llama-3.3-70b-versatile, etc.)
• Fireworks (fireworks/accounts/fireworks/models/llama-v3p3-70b-instruct, etc.)
• OpenRouter (openrouter/openai/gpt-4o, etc.)
• Vercel (vercel/openai/gpt-4o, etc.)
• Superagent (superagent/guard-1.7b, etc.) - Default for guard

Environment Variables

Configure provider API keys:

export SUPERAGENT_API_KEY=your-superagent-key
export OPENAI_API_KEY=your-openai-key
export OPENAI_COMPATIBLE_API_KEY=your-openai-compatible-key
export OPENAI_COMPATIBLE_BASE_URL=https://your-endpoint/v1
export ANTHROPIC_API_KEY=your-anthropic-key
export GOOGLE_API_KEY=your-google-key
export GROQ_API_KEY=your-groq-key
export FIREWORKS_API_KEY=your-fireworks-key
export OPENROUTER_API_KEY=your-openrouter-key
export AI_GATEWAY_API_KEY=your-vercel-key

License

MIT