Responsible AI Overview

Learn about responsible AI practices, safety controls, and governance tools in Microsoft Foundry for building trustworthy AI agents.

Responsible AI for Microsoft Foundry

Microsoft Foundry provides comprehensive tools and practices for building trustworthy AI agents with end-to-end security, observability, and governance.

Responsible AI Framework

Our approach follows the Microsoft Responsible AI Standard with three key stages:

1. Discover

   Identify agent quality, safety, and security risks before and after deployment

2. Protect

   Apply safeguards at model output and agent runtime levels

3. Govern

   Monitor agents through tracing, logging, and compliance integrations

Safety Controls

Content Filtering

Capabilities:

• Block harmful inputs and outputs
• Detect and prevent jailbreak attempts
• Mitigate prompt injection attacks (including XPIA)
• Filter for multiple harm categories

Categories:

• Violence and self-harm
• Sexual content
• Hate speech
• Protected material

Prompt Injection Protection

Protections:

• Input sanitization
• Context isolation
• Cross-prompt injection attack (XPIA) detection
• User/system message separation

Observability

Conversation Tracing

Features:

• Full conversation history
• Tool invocation logs
• Decision point tracking
• Performance metrics

Application Insights Integration

Capabilities:

• Usage data and metrics
• Performance monitoring
• Error tracking
• Custom telemetry

Evaluation Tools

Assessment types:

• Quality evaluations
• Safety assessments
• Performance benchmarks
• Custom metrics

Governance

Identity and Access

Controls:

• Microsoft Entra ID authentication
• Role-based access control (RBAC)
• Managed identities
• Conditional access policies

Data Governance

Capabilities:

• Data residency controls
• Customer-managed encryption keys
• Audit logging
• Compliance certifications

Network Security

Features:

• Private endpoints
• Virtual network integration
• Network isolation
• Data exfiltration prevention

Security Alerts

View and respond to security alerts in the Risks + alerts section:

1. Sign in

   Navigate to Microsoft Foundry portal

2. Select Project

   Choose your project

3. View Alerts

   Go to Risks + alerts in left navigation

4. Review & Act

   Review active alerts and take remediation steps

Best Practices

Test Before Deployment

• Run adversarial testing
• Evaluate with representative data
• Test edge cases and failure modes
• Conduct red team exercises

Monitor Continuously

• Set up Application Insights
• Track key performance metrics
• Monitor for anomalies
• Review conversation logs regularly

Implement Defense in Depth

• Use content filters
• Enable prompt injection protection
• Apply RBAC and network controls
• Encrypt data at rest and in transit

Document and Audit

• Maintain agent documentation
• Log all configurations
• Review audit logs
• Document incident responses

Resources

• Microsoft AI Principles
• Responsible AI Resources
• Azure AI Content Safety
• Microsoft Responsible AI Standard

Next Steps

Agent Overview

Learn about agent capabilities

Standard Setup

Configure enterprise security

Environment Setup

Deploy secure infrastructure

Quickstart

Build your first agent