agentshield 0.1.0

Security and monitoring for AI agents

AgentShield Python SDK

Security and monitoring for AI agents. Monitor agent behavior, enforce policies, and get real-time alerts for suspicious activity.

AgentShield provides enterprise-grade security for autonomous AI agents, with seamless integration for LangChain, OpenAI Assistants, and custom agent frameworks.

🔗 Dashboard: https://agent-shield.com 📚 Documentation: https://agent-shield.com/docs

---

✨ Features

• 🛡️ Policy Enforcement: Block, flag, or require approval for sensitive operations
• 📊 Real-Time Monitoring: Track every tool call, API request, and agent decision
• 🚨 Smart Alerts: Get notified of anomalies, policy violations, and security events
• 🔍 Anomaly Detection: ML-powered detection of unusual agent behavior
• ⚡ Zero-Config: Drop-in replacement - no code changes required
• 🎯 Framework Agnostic: Works with LangChain, OpenAI, or custom agents
• 📈 Cost Tracking: Monitor and optimize agent operational costs
• 🔐 Fail-Safe: Configurable fail-open or fail-closed modes

---

🚀 Quick Start

Installation

pip install agentshield

5-Line Integration

from agentshield import SecureAgent
from langchain.agents import create_openai_functions_agent

# Your existing agent
agent = create_openai_functions_agent(llm, tools, prompt)

# Add security - that's it!
secure_agent = SecureAgent(
    agent=agent,
    shield_key="agsh_your_api_key",
    agent_id="my-assistant"
)

# Use normally - security is automatic
result = secure_agent.invoke({"input": "Search for AI security"})

Get your API key: https://agent-shield.com/dashboard/settings

---

📖 Examples

Basic Function Wrapping

Secure any Python function with AgentShield:

from agentshield import SecureAgent, SecurityException

def web_search(query: str) -> str:
    """Search the web."""
    return perform_search(query)

# Initialize SecureAgent
secure_agent = SecureAgent(
    agent=agent,
    shield_key="agsh_...",
    agent_id="search-agent"
)

# Wrap the function
secure_search = secure_agent.wrap_function(web_search, "web_search")

# Use it - calls are automatically monitored
try:
    result = secure_search("AI security best practices")
    print(result)
except SecurityException as e:
    print(f"Blocked: {e.message}")
    print(f"Policy: {e.policy_matched}")
    print(f"Call ID: {e.call_id}")

LangChain Integration

AgentShield automatically wraps LangChain tools:

from langchain.agents import AgentExecutor, create_openai_functions_agent
from langchain.tools import Tool
from langchain_openai import ChatOpenAI
from agentshield import SecureAgent

# Define tools
tools = [
    Tool(name="Search", func=search_func, description="Search the web"),
    Tool(name="Calculator", func=calc_func, description="Do math"),
    Tool(name="Database", func=db_func, description="Query database"),
]

# Create LangChain agent
llm = ChatOpenAI(model="gpt-4")
agent = create_openai_functions_agent(llm, tools, prompt)
agent_executor = AgentExecutor(agent=agent, tools=tools, verbose=True)

# Wrap with AgentShield - tools are automatically secured
secure_agent = SecureAgent(
    agent=agent_executor,
    shield_key="agsh_...",
    agent_id="langchain-assistant",
    fail_open=False,  # Fail closed for security
)

# Use normally - all tool calls are monitored and enforced
result = secure_agent.invoke({"input": "What's the weather in SF?"})

OpenAI Assistants

from openai import OpenAI
from agentshield import SecureAgent

client = OpenAI()

# Create assistant
assistant = client.beta.assistants.create(
    name="Data Analyst",
    instructions="You analyze data and generate reports.",
    tools=[{"type": "code_interpreter"}],
    model="gpt-4"
)

# Wrap with AgentShield
secure_assistant = SecureAgent(
    agent=assistant,
    shield_key="agsh_...",
    agent_id="data-analyst",
)

# Function calls are now monitored
thread = client.beta.threads.create()
message = client.beta.threads.messages.create(
    thread_id=thread.id,
    role="user",
    content="Analyze this dataset and find trends"
)
run = client.beta.threads.runs.create(
    thread_id=thread.id,
    assistant_id=secure_assistant.id
)

---

🎛️ Configuration

SecureAgent Options

secure_agent = SecureAgent(
    agent=agent,                    # Required: Your agent instance
    shield_key="agsh_...",          # Required: Your AgentShield API key
    agent_id="unique-agent-id",     # Required: Unique identifier for this agent
    api_url=None,                   # Optional: Custom API endpoint
    timeout=30,                     # Optional: API timeout in seconds
    debug=False,                    # Optional: Enable debug logging
    fail_open=False,                # Optional: Allow calls if API unavailable
)

Policy Actions

AgentShield supports four policy actions:

Action	Behavior	Use Case
ALLOWED	Execute tool normally	Safe, approved operations
BLOCKED	Raise SecurityException, prevent execution	Dangerous operations, PII access
FLAGGED	Log warning, allow execution	Suspicious but not critical
PENDING_APPROVAL	Raise exception, require manual approval	High-risk operations

---

🔐 Security Policies

Create policies in the AgentShield Dashboard:

Example Policies

Block PII Access

name: Block PII Access
action: BLOCK
conditions:
  keywords: ["social security", "ssn", "credit card", "password"]
  tools: ["database_query", "file_read"]

Flag Expensive Operations

name: Flag Expensive Operations
action: FLAG
conditions:
  cost_limit: 1.00  # Flag if cost > $1
  tools: ["gpt4_call", "dalle_generate"]

Rate Limiting

name: Rate Limit API Calls
action: BLOCK
conditions:
  rate_limit:
    calls: 100
    window: 3600  # 100 calls per hour

---

📊 Monitoring & Alerts

Dashboard Features

Visit agent-shield.com/dashboard to:

• 📈 View Activity: See every agent call in real-time
• 🎯 Create Policies: Define security rules with a visual editor
• 🚨 Set Alerts: Get notified via email or webhook
• 📉 Track Metrics: Monitor costs, performance, and anomalies
• 🔍 Search Logs: Find specific calls, filter by status/tool/agent
• 📊 Analytics: Understand agent behavior patterns

Webhook Alerts

Configure webhooks to receive alerts:

# Set webhook URL in dashboard settings
webhook_url = "https://your-api.com/agentshield-webhook"

# Receive events:
{
  "event": "call_blocked",
  "agent_id": "my-agent",
  "tool_name": "database_query",
  "policy_matched": "Block PII Access",
  "call_id": "call_abc123",
  "timestamp": "2024-11-06T12:34:56Z",
  "severity": "HIGH"
}

---

🛠️ Error Handling

SecurityException

from agentshield import SecurityException

try:
    result = secure_agent.invoke({"input": "Delete user data"})
except SecurityException as e:
    print(f"Status: {e.status}")                # BLOCKED or PENDING_APPROVAL
    print(f"Message: {e.message}")              # Human-readable reason
    print(f"Policy: {e.policy_matched}")        # Which policy was violated
    print(f"Call ID: {e.call_id}")              # For debugging/auditing

Other Exceptions

from agentshield import (
    APIKeyError,          # Invalid or revoked API key
    NetworkError,         # Network/connectivity issues
    PolicyEvaluationError,  # Server-side policy evaluation failed
    ConfigurationError,   # Invalid SDK configuration
)

Fail-Open vs Fail-Closed

# Fail-Closed (default): Block calls if API is unavailable
secure_agent = SecureAgent(agent, shield_key, agent_id, fail_open=False)

# Fail-Open: Allow calls if API is unavailable (less secure)
secure_agent = SecureAgent(agent, shield_key, agent_id, fail_open=True)

---

🧪 Testing

# Install dev dependencies
pip install -e ".[dev]"

# Run tests
pytest

# Run with coverage
pytest --cov=agentshield --cov-report=html

# Run specific test
pytest tests/test_client.py::TestAgentShieldClient::test_log_agent_call_allowed

---

📚 API Reference

SecureAgent

__init__(agent, shield_key, agent_id, **options)

Initialize SecureAgent wrapper.

wrap_function(func, tool_name=None) -> Callable

Manually wrap a function with AgentShield security.

AgentShieldClient

log_agent_call(tool_name, tool_args, execution_time_ms=None, timestamp=None) -> dict

Log an agent call to AgentShield for policy evaluation.

---

🔧 Troubleshooting

Common Issues

1. APIKeyError: Invalid or revoked shield_key

• Check your API key at agent-shield.com/dashboard/settings
• Ensure key starts with agsh_
• Key may be revoked - generate a new one

2. NetworkError: Failed to connect

• Check your internet connection
• Verify firewall allows outbound HTTPS to *.cloudfunctions.net
• Try with timeout=60 for slower connections

3. Tool calls not being intercepted

• Enable debug mode: SecureAgent(..., debug=True)
• Check logs for "Wrapped X tools" message
• Some agent frameworks require manual wrapping

4. High latency

• API calls add ~50-200ms overhead
• Use fail_open=True for non-critical applications
• Consider caching for repeated calls

Debug Mode

import logging

# Enable debug logging
logging.basicConfig(level=logging.DEBUG)

secure_agent = SecureAgent(
    agent=agent,
    shield_key="agsh_...",
    agent_id="debug-agent",
    debug=True  # Enables verbose logging
)

---

🤝 Contributing

We welcome contributions! Please see our Contributing Guide.

# Setup development environment
git clone https://github.com/agentshield/python-sdk.git
cd python-sdk
pip install -e ".[dev]"

# Run tests
pytest

# Format code
black agentshield tests examples
flake8 agentshield

# Type check
mypy agentshield

---

📄 License

MIT License - see LICENSE file for details.

---

🔗 Links

• Dashboard: https://agent-shield.com
• Documentation: https://agent-shield.com/docs
• GitHub: https://github.com/agentshield/python-sdk
• PyPI: https://pypi.org/project/agentshield/
• Support: [e-mail]

---

🙋 Support

• Email: [e-mail]
• GitHub Issues: github.com/agentshield/python-sdk/issues
• Documentation: agent-shield.com/docs
• Discord: discord.gg/agentshield

---

Made with ❤️ by the AgentShield team