GENSEC03-BP01 Implement control plane and data access monitoring to generative AI services and foundation models
Implement comprehensive monitoring across both control and data planes to enhance the protection of generative AI workloads against service-level misconfigurations. This monitoring and auditing approach enables tracking of key aspects such as application performance, workload quality, and security.
Desired outcome: When implemented, you can track the changes made to generative AI services and infrastructure, as well as changes to relevant data stores.
Benefits of establishing this best practice: Apply security at all layers - Control and data plane monitoring provides a layer of security at the service configuration and data access layers.
Level of risk exposed if this best practice is not established: High
Implementation guidance
Monitoring at the control plane and data layers should track data access, as well as control plane API requests to the services in question. Most cloud-based systems publish these events over an event bus for capture, storage, and eventual analysis. These capabilities are considered normal within a modern data architecture. As data and AI workloads become more closely intertwined in your organization, solutions like Amazon SageMaker AI and its new Lakehouse capability help simplify the collection and capturing of data access requests by models, workloads, and users. Your organization AI policy document should define how data access requests are captured and monitored across your environment.
Consider AWS CloudTrail to record management and data events. Amazon Bedrock, Amazon Q Business, and other generative AI services integrate with CloudTrail and can be used to record control plane operations like custom model import and runtime operations like invokeAgent. Amazon CloudWatch can be configured to capture logs for generative AI applications as well. A combination of these AWS services or the use of a third-party logging solution, if needed, improves visibility into application security. CloudWatch and CloudTrail integrate well with other managed AWS services powered by data, such Quick Q, a generative business intelligence (BI) tool.
Implementation steps
-
Performance monitoring:
-
Track response times, latency, and throughput of model inference
-
Monitor resource utilization (CPU, GPU, and memory)
-
Measure token usage and request volumes
-
Track batch processing efficiency and queue lengths
-
Monitor model loading and unloading times
-
-
Quality and accuracy monitoring:
-
Track completion rates and success ratios
-
Monitor response quality scores
-
Implement content safety measurements
-
Track hallucination rates and accuracy metrics
-
Monitor prompt effectiveness and completion relevance
-
-
Security monitoring:
-
Track authentication and authorization attempts
-
Monitor for potential prompt injection exploits
-
Log access patterns and unusual behaviors
-
Track rate limiting and quota usage
-
Monitor for potential data leakage
-
-
Cost monitoring:
-
Track token usage and associated costs
-
Monitor resource utilization costs
-
Track API call volumes and expenses
-
Monitor storage and data transfer costs
-
Track model deployment and training costs
-
-
Audit trail implementation:
-
Maintain detailed logs of requests and responses
-
Record user interactions and system changes
-
Log model version changes and updates
-
Track configuration modifications
-
Maintain compliance-related audit trails
-
-
Compliance monitoring:
-
Track data retention compliance
-
Monitor PII handling and protection
-
Verify regulatory requirement adherence
-
Track consent management
-
Monitor geographic data restrictions
-
Resources
Related best practices:
Related documents:
Related examples:
