Generative AI protection in AWS

AWS has introduced several new features to enhance the protection of generative AI workloads:

1. Amazon Bedrock Guardrails: This feature allows you to apply safeguards across multiple foundation models, knowledge bases, and agents. It helps filter harmful content, block denied topics, and redact sensitive information.

2. Secure Prompt Engineering: AWS recommends using prompt templates and system prompts to structure inputs and constrain model behavior, limiting the ability of malicious users to manipulate core instructions.

3. Enhanced Monitoring and Logging: AWS provides services like AWS CloudTrail and Amazon Bedrock model invocation logs to help monitor and log your generative AI application. Amazon Bedrock Guardrails also offers tracing capabilities to provide insights into how the guardrail evaluated inputs and outputs.

4. Model Evaluation on Amazon Bedrock: This feature helps customers evaluate, compare, and select the best foundation models for their specific use cases based on custom metrics.

5. Watermark Detection for Amazon Titan Image Generator: This allows users to identify images generated by the Amazon Titan Image Generator, helping to mitigate harmful content generation and reduce the spread of misinformation.

6. Encrypted Training Data and Isolated Environments: When customizing models, Amazon Bedrock uses encrypted training data and deploys fine-tuned models in isolated environments for enhanced security.

These features are part of AWS's ongoing efforts to provide a secure, compliant, and responsible foundation for building generative AI applications.
Sources
Safeguard your generative AI workloads from prompt injections | AWS Security Blog
Build generative AI applications on Amazon Bedrock — the secure, compliant, and responsible foundation | AWS Machine Learning Blog