AWS wellarchitected high security documentation change
Summary
Added two new best practices: FSIOPS02-BP03 (Evaluate data privacy/security for generative AI) and FSIOPS02-BP04 (Establish prompt engineering standards).
Security assessment
Directly addresses security vulnerabilities in generative AI including data leakage prevention (tokenization, KMS encryption), prompt injection attacks, and unauthorized data exposure. Specific evidence includes implementation of Guardrails, KMS encryption, and explicit mention of prompt injection attacks.
Diff
diff --git a/wellarchitected/latest/financial-services-industry-lens/fsiops2.md b/wellarchitected/latest/financial-services-industry-lens/fsiops2.md index 8387cc997..f14bce93a 100644 --- a//wellarchitected/latest/financial-services-industry-lens/fsiops2.md +++ b//wellarchitected/latest/financial-services-industry-lens/fsiops2.md @@ -5 +5 @@ -FSIOPS02-BP01 Understand the Shared Responsibility Model and how it applies to services and workloads you run in the cloudFSIOPS02-BP02 Develop an enterprise cloud risk plan +FSIOPS02-BP01 Understand the Shared Responsibility Model and how it applies to services and workloads you run in the cloudFSIOPS02-BP02 Develop an enterprise cloud risk planFSIOPS02-BP03 Evaluate data privacy and security requirements for generative AIFSIOPS02-BP04 Establish prompt engineering standards and version control @@ -28,0 +29,28 @@ Evaluate existing risk models in use, and related policies, for relevance in a c +## FSIOPS02-BP03 Evaluate data privacy and security requirements for generative AI + +Generative AI models require careful consideration of data handling, especially when processing sensitive financial information. Implement data classification, tokenization, and privacy-preserving techniques when using foundation models. Adhere to data residency requirements and understand the data processing practices of third-party model providers. Establish data retention policies and ensure generative AI systems support regulatory requirements including data subject rights. + +### Prescriptive guidance + +Use Amazon Bedrock with AWS PrivateLink to implement network isolation for generative AI inference. Implement data masking and tokenization before sending sensitive data to foundation models. Configure Amazon Bedrock Guardrails to prevent unauthorized data exposure in model outputs. Use AWS KMS for encryption of prompts and responses containing sensitive information. Document data flow diagrams showing how sensitive data moves through generative AI pipelines including retention periods and deletion schedules. + +Implement AWS CloudTrail and Amazon CloudWatch for comprehensive audit logging of data access and model interactions. + +Define specific data retention periods for prompts, responses, and model training data in accordance with regulatory requirements. + +## FSIOPS02-BP04 Establish prompt engineering standards and version control + +Prompts are critical operational assets in generative AI systems requiring comprehensive governance frameworks. Implement version control, testing, and approval processes for prompt templates used in production. Establish prompt engineering best practices and security guidelines to prevent prompt injection attacks. + +### Prescriptive guidance + +Store production prompts in AWS CodeCommit or similar version control systems with change tracking and approval workflows. + +Implement automated testing for prompt templates using representative test cases. Use AWS Lambda and AWS Step Functions to create controlled prompt execution pipelines with automated rollback capabilities. + +Establish prompt security guidelines including input validation, sanitization and protection against prompt injection attacks. + +Establish prompt performance monitoring to track effectiveness and model response quality over time with automated alerting. + +Define escalation procedures for prompt-related security incidents and integrate with existing incident response frameworks. +