AWS wellarchitected medium security documentation change
Summary
Added two new best practices: FSIOPS07-BP03 (feedback loops for model improvement) and FSIOPS07-BP04 (generative AI-specific chaos engineering) with implementation guidance
Security assessment
The change adds explicit security-focused content: 1) FSIOPS07-BP03 includes bias detection and fairness analysis using Amazon SageMaker AI Clarify 2) FSIOPS07-BP04 includes testing for bias amplification scenarios and security mechanisms like fallback procedures and human oversight validation during failures
Diff
diff --git a/wellarchitected/latest/financial-services-industry-lens/fsiops7.md b/wellarchitected/latest/financial-services-industry-lens/fsiops7.md index da86ad02a..8c94f7aa5 100644 --- a//wellarchitected/latest/financial-services-industry-lens/fsiops7.md +++ b//wellarchitected/latest/financial-services-industry-lens/fsiops7.md @@ -5 +5 @@ -FSIOPS07-BP01 Test, model, and simulate scenarios before rolloutFSIOPS07-BP02 Conduct post-event operational reviews +FSIOPS07-BP01 Test, model, and simulate scenarios before rolloutFSIOPS07-BP02 Conduct post-event operational reviewsFSIOPS07-BP03 Implement feedback loops for model improvement FSIOPS07-BP04 Conduct generative AI-specific chaos engineering @@ -26,0 +27,28 @@ Post-event operational reviews are comprised of two components: identification o +## FSIOPS07-BP03 Implement feedback loops for model improvement + +Establish mechanisms to capture user feedback on generative AI outputs and use this data to improve prompt engineering, model selection, bias detection, and operational procedures. Create processes for incorporating lessons learned into model governance and operational practices. + +### Prescriptive guidance + +Deploy feedback collection mechanisms using Amazon DynamoDB to store user ratings and comments. Use Amazon Comprehend to analyze feedback sentiment and identify improvement areas. Implement A/B testing frameworks using AWS Lambda to compare different models or prompts. Create monthly operational reviews focused on generative AI metrics and improvement opportunities. Use Amazon SageMaker AI Clarify for automated bias detection and fairness analysis based on feedback patterns. Implement Amazon Athena for advanced analytics on feedback trends and correlation analysis. + +## FSIOPS07-BP04 Conduct generative AI-specific chaos engineering + +Test the [resilience of generative AI workloads](https://catalog.us-east-1.prod.workshops.aws/workshops/d56fd754-5e56-43c5-addc-d69ac130a099/en-US) through controlled experiments including model API failures, rate limiting scenarios, quality degradation simulations, and bias amplification scenarios. Validate that fallback mechanisms and human oversight processes function correctly under stress. + +### Prescriptive guidance + +Use [AWS Fault Injection Service](https://builder.aws.com/content/2uSMnBJb3h7JxB9SkryFvXfQWk8/chaos-engineering-scenarios-for-genai-workloads) to test generative AI workload resilience. + +Simulate model API throttling, timeout scenarios, and complete service or model unavailability to test failover mechanisms and business continuity procedures. + +Test fallback mechanisms when primary models are unavailable including automated switching to backup models. + +Validate that human review processes can handle increased load during model failures. + +Test system behavior when input data quality deteriorates to ensure graceful degradation and appropriate human intervention triggers. + +Simulate bias amplification scenarios to test detection mechanisms and response procedures for maintaining fair and compliant AI outputs. + +Test cross-system dependencies by simulating failures in databases, APIs, and other generative AI services' dependent systems. +