AI Agent Security

Bastio provides comprehensive security for AI agents that use tools. As AI agents become more autonomous and capable, they gain access to powerful capabilities - executing shell commands, accessing files, making HTTP requests, and interacting with databases. Without proper guardrails, these capabilities become security vulnerabilities.

The Challenge

Modern AI agents face several security challenges:

• Autonomous Tool Execution: Agents can execute dangerous operations without human oversight
• Prompt Injection Attacks: Malicious inputs can hijack agent behavior and tool usage
• Data Exfiltration Risks: Agents with network access can leak sensitive information
• Privilege Escalation: Tool chains can be exploited to gain unauthorized access
• Compliance Requirements: Regulated industries need audit trails and approval workflows

How Bastio Protects Your Agents

Bastio provides a three-layer security model for AI agent tool calls:

Layer 1: Policy Engine (Primary Enforcement)

Define rules for how tool calls should be handled:

• Allow: Permit tool execution
• Block: Prevent tool execution
• Require Approval: Route to human reviewers
• Rate Limit: Prevent abuse through request throttling
• Sanitize: Clean dangerous content from arguments
• Warn: Allow but flag for monitoring

Policies can match tools by name pattern, evaluate argument content, consider risk scores, and apply time-based conditions.

Layer 2: Real-Time Threat Scanning

Every tool call is scanned in real-time for threats:

• Shell Injection: Commands like ; rm -rf / or | cat /etc/passwd
• File Access Attacks: Attempts to read sensitive files (~/.ssh, .env)
• Network Abuse: Data exfiltration URLs, reverse shells
• Prompt Injection: Attempts to manipulate agent behavior
• Privilege Escalation: sudo, chmod 777, admin tool calls
• Credential Exposure: API keys, passwords, tokens in arguments

Risk scoring (0.0 - 1.0) influences policy decisions.

Layer 3: Agent Identity & Authentication

Cryptographic identity for your AI agents:

• Ed25519 Key Pairs: Strong cryptographic authentication
• Trust Levels: Differentiate development vs production agents
• Session Management: Track agent activity over time
• Key Rotation: Security best practice support
• Allowed/Blocked Tools: Per-agent tool restrictions

Quick Start

1. Get Your API Key

Create a Gateway API key from your Bastio Dashboard.

2. Validate Tool Calls

Before executing any tool, send it to Bastio for validation:

curl -X POST https://api.bastio.com/v1/guard/{proxyId}/agent/validate \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "session_id": "session_123",
    "tool_calls": [
      {
        "id": "call_abc",
        "name": "execute_shell",
        "type": "function",
        "arguments": {"command": "ls -la"}
      }
    ]
  }'

3. Handle the Response

{
  "allowed": true,
  "tool_results": [
    {
      "tool_id": "call_abc",
      "allowed": true,
      "action": "allowed",
      "risk_score": 0.15,
      "threats": []
    }
  ],
  "blocked_count": 0,
  "risk_score": 0.15,
  "request_id": "req_xyz789"
}

Possible actions per tool:

• allowed - Safe to execute the tool
• blocked - Do not execute, return error to agent
• pending_approval - Wait for human approval before executing
• warned - Execute but flag for review
• sanitized - Execute with cleaned arguments

Key Capabilities

Supported Formats

Bastio supports multiple tool call formats:

{
  "tool_calls": [
    {
      "id": "call_abc123",
      "name": "get_weather",
      "type": "function",
      "arguments": {"location": "San Francisco"}
    }
  ]
}

{
  "tool_calls": [
    {
      "id": "call_abc123",
      "type": "function",
      "function": {
        "name": "get_weather",
        "arguments": "{\"location\": \"San Francisco\"}"
      }
    }
  ]
}

{
  "tool_use": [
    {
      "type": "tool_use",
      "id": "toolu_01abc",
      "name": "get_weather",
      "input": {"location": "San Francisco"}
    }
  ]
}

SDK (Beta)

We're developing official SDKs to make integration even easier:

• TypeScript SDK: @bastio/agent-security (coming soon)
• Python SDK: bastio-agent-security (coming soon)

Beta testers receive direct engineering support and free Pro tier access during the beta period.

Next Steps

• Tool Validation - Learn the complete validation API
• Policies - Configure custom security rules
• Human-in-the-Loop - Set up approval workflows
• Agent Identity - Register and authenticate agents