# FSISEC14: How do you monitor AI system outputs for security issues?
<a name="fsisec14"></a>

 Continuous monitoring of AI system outputs is critical for financial institutions to detect harmful responses, potential data leakage, and security violations. Without proper monitoring, AI systems may generate responses that expose sensitive information, violate compliance requirements, or create security vulnerabilities. Implementing comprehensive monitoring across all AI interactions enables organizations to identify and address security issues before they impact customers or operations. 

## FSISEC14-BP01 Implement automated response validation
<a name="fsisec14-bp01-implement-automated-response-validation"></a>

 Automated response validation is essential for ensuring AI systems operate within defined security parameters. Deploy guardrails for content filtering to detect and prevent harmful, biased, or non-compliant responses from reaching users. Monitor for prompt injection attempts where malicious inputs might manipulate model behavior and implement automated detection systems that flag potentially harmful responses for review. 

 Establish clear response quality and safety metrics that align with your organization's security and compliance requirements. Create alert mechanisms that notify security teams when suspicious AI system behavior is detected, enabling rapid investigation and remediation of potential security issues. 

## FSISEC14-BP02 Monitor AI system interactions
<a name="fsisec14-bp02-monitor-ai-system-interactions"></a>

 Comprehensive monitoring of AI system interactions provides visibility into potential security issues and enables proactive threat detection. Track all model invocations and user interactions to establish usage patterns and identify anomalies that may indicate security incidents. Monitor for unauthorized access patterns to AI services that could signal credential compromise or insider threats. 

 Implement comprehensive logging of AI system events including user inputs, model responses, and system actions. Establish baseline behavior patterns for AI systems to enable anomaly detection and monitor for potential data leakage in model responses that could expose sensitive financial information or intellectual property. 

## FSISEC14-BP03 Establish AI incident response procedures
<a name="fsisec14-bp03-establish-ai-incident-response-procedures"></a>

 Financial institutions must develop specialized incident response procedures for AI-specific security events. Develop playbooks that address unique AI security incidents such as prompt injection attacks, harmful model responses, or model manipulation attempts. Include harmful model responses in your incident classification system to ensure appropriate escalation and response. 

 Establish clear procedures for handling model response validation failures, including containment, investigation, and remediation steps. Create escalation procedures for AI security events that define roles, responsibilities, and communication channels. Where appropriate, implement automated response mechanisms that can take immediate action when AI security issues are detected, such as blocking suspicious requests or disabling compromised endpoints. 

## Resources
<a name="resources-2"></a>

### Documents
<a name="documents-1"></a>
+  [AWS Well-Architected Generative AI Lens - Governance](https://docs.aws.amazon.com/wellarchitected/latest/generative-ai-lens/governance.html) 
+  [IAM Best Practices for AI Services](https://docs.aws.amazon.com/IAM/latest/UserGuide/best-practices.html) 
+  [Amazon SageMaker AI Model Governance](https://docs.aws.amazon.com/sagemaker/latest/dg/model-governance.html)