Security

​

Overview

​

Security Architecture

​

Credentials Proxy

1. Agent makes API call through Envoy proxy
2. Envoy intercepts the request
3. Credential Resolver fetches secrets from vault
4. Envoy injects credentials at request time
5. Request is forwarded to external API
6. Secrets are never stored in agent memory

​

Benefits

​

Claude Sandbox Isolation

​

gVisor Isolation

• User-space kernel intercepts all syscalls
• Reduced kernel attack surface
• Container-to-host isolation

​

Network Policies

apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
  name: agent-sandbox
spec:
  podSelector:
    matchLabels:
      app: sre-agent
  policyTypes:
    - Egress
  egress:
    - to:
        - namespaceSelector:
            matchLabels:
              name: incidentfox
    - ports:
        - port: 443  # HTTPS only

​

Resource Limits

• CPU: Bounded to prevent runaway processes
• Memory: Capped to prevent OOM attacks
• Time: Maximum investigation duration
• Ephemeral: Sandbox destroyed after use

​

Authentication

​

Token Types

​

SSO/OIDC Support

• Google Workspace
• Azure AD / Entra ID
• Okta
• Generic OIDC

​

Configuration

{
  "auth": {
    "oidc": {
      "issuer": "https://accounts.google.com",
      "client_id": "your-client-id",
      "allowed_domains": ["your-company.com"]
    }
  }
}

​

Authorization (RBAC)

​

Roles

​

Tool-Level Permissions

{
  "tools": {
    "docker_exec": {
      "enabled": true,
      "allowed_roles": ["admin"],
      "require_approval": true
    }
  }
}

​

Approval Workflows

{
  "approval_workflows": {
    "enabled": true,
    "actions": {
      "pod_restart": {
        "required_approvers": 1,
        "timeout_minutes": 30,
        "notify_channel": "#sre-approvals"
      },
      "scale_deployment": {
        "required_approvers": 2,
        "timeout_minutes": 15
      }
    }
  }
}

​

Approval Flow

1. Agent proposes action
2. Notification sent to approvers
3. Approver reviews and approves/denies
4. Action executed or cancelled

​

Audit Logging

​

Event Types

​

Log Format

{
  "timestamp": "2024-01-15T14:30:00Z",
  "event_type": "tool_executed",
  "user_id": "user@company.com",
  "team_id": "team_123",
  "tool": "get_pod_logs",
  "parameters": {
    "namespace": "production",
    "pod": "api-server-xyz"
  },
  "duration_ms": 1234,
  "status": "success"
}

​

Log Destinations

• CloudWatch Logs
• Datadog
• Splunk
• Custom webhook

​

Compliance

​

SOC 2 Type II

​

Data Handling

​

Deployment Security

​

Self-Hosted

• Deploy in your VPC
• Use your secrets manager
• Control all network egress
• Manage your own keys

​

Air-Gapped

• No external network access
• Local model inference
• Internal secrets management
• Manual updates

​

Best Practices

​

Credential Management

1. Use vault references, never plain text
2. Rotate credentials regularly
3. Use service accounts with minimal permissions
4. Enable audit logging for secret access

​

Network Security

1. Use private endpoints where possible
2. Enable VPC peering for cloud services
3. Restrict agent egress to necessary destinations
4. Use TLS for all communications

​

Access Control

1. Enable SSO for all users
2. Use team-scoped tokens
3. Require MFA for admin access
4. Regular access reviews

​

Monitoring

1. Alert on authentication failures
2. Monitor for unusual tool usage
3. Track investigation patterns
4. Review audit logs regularly

​

Incident Response

​

Security Incidents

1. Email security@incidentfox.ai
2. Do not disclose publicly
3. We will respond within 24 hours

​

Vulnerability Disclosure

• 90-day disclosure timeline
• Credit for reporters
• Bug bounty program available

​

Next Steps