AWS generative AI best practices framework v1 - AWS Audit Manager

AWS generative AI best practices framework v1

AWS Audit Manager provides a prebuilt standard framework to help you gain visibility into how your generative AI implementation on Amazon Bedrock is working against AWS recommended best practices.

Amazon Bedrock is a fully managed service that makes AI models from Amazon and other leading AI companies available through an API. With Amazon Bedrock, you can privately tune existing models with your organization’s data. This enables you to harness foundation models (FMs) and large language models (LLMs) to build applications securely, without compromising data privacy. For more information, see What is Amazon Bedrock? in the Amazon Bedrock User Guide.

What are AWS generative AI best practices for Amazon Bedrock?

Generative AI refers to a branch of AI that focuses on enabling machines to generate content. Generative AI models are designed to create outputs that closely resemble the examples that they were trained on. This creates scenarios where AI can mimic human conversation, generate creative content, analyze vast volumes of data, and automate processes that are normally done by humans. The rapid growth of generative AI brings promising new innovation. At the same time, it raises new challenges around how to use generative AI responsibly and in compliance with governance requirements.

AWS is committed to providing you with the tools and guidance needed to build and govern applications responsibly. To help you with this goal, Audit Manager has partnered with Amazon Bedrock to create the AWS generative AI best practices framework v1. This framework provides you with a purpose-built tool for monitoring and improving the governance of your generative AI projects on Amazon Bedrock. You can use the best practices in this framework to gain tighter control and visibility over your model usage and stay informed on model behavior.

The controls in this framework were developed in collaboration with AI experts, compliance practitioners, security assurance specialists across AWS, and with input from Deloitte. Each automated control maps to an AWS data source from which Audit Manager collects evidence. You can use the collected evidence to evaluate your generative AI implementation based on the following eight principles:

  1. Responsible – Develop and adhere to ethical guidelines for the deployment and usage of generative AI models

  2. Safe – Establish clear parameters and ethical boundaries to prevent the generation of harmful or problematic output

  3. Fair – Consider and respect how an AI system impacts different sub-populations of users

  4. Sustainable – Strive for greater efficiency and more sustainable power sources

  5. Resilience – Maintain integrity and availability mechanisms to ensure an AI system operates reliably

  6. Privacy – Ensure that sensitive data is protected from theft and exposure

  7. Accuracy – Build AI systems that are accurate, reliable, and robust

  8. Secure – Prevent unauthorized access to generative AI systems

Example

Let's say that your application uses a third-party foundational model that’s available on Amazon Bedrock. You can use the AWS generative AI best practices framework to monitor your usage of this model. By using this framework, you can collect evidence that demonstrates that your usage is compliant with generative AI best practices. This provides you with a consistent approach for tracking track model usage and permissions, flagging sensitive data, and being alerted about any inadvertent disclosures. For instance, specific controls in this framework can collect evidence that helps you show that you’ve implemented mechanisms for the following:

  • Documenting the source, nature, quality, and treatment of the new data, to ensure transparency and help in troubleshooting or audits (Responsible)

  • Regularly evaluating the model using predefined performance metrics to ensure it meets accuracy and safety benchmarks (Safe)

  • Using automated monitoring tools to detect and alert on potential biased outcomes or behaviors in real-time (Fair)

  • Evaluating, identifying, and documenting model usage and scenarios where existing models can be reused, whether you generated them or not (Sustainable)

  • Setting up procedures for notification if there is inadvertent PII spillage or unintentional disclosure (Privacy)

  • Establishing real-time monitoring of the AI system and setting up alerts for any anomalies or disruptions (Resilience)

  • Detecting inaccuracies, and conducting a thorough error analysis to understand the root causes (Accuracy)

  • Implementing end-to-end encryption for input and output data of the AI models to minimum industry standards (Secure)

Using this framework to support your audit preparation

Note
  • If you're an Amazon Bedrock customer, you can use this framework directly in Audit Manager. Make sure that you use the framework and run assessments in the AWS accounts and Regions where you run your generative AI models and applications.

  • If you want to encrypt your CloudWatch logs for Amazon Bedrock with your own KMS key, make sure that Audit Manager has access to that key. To do this, you can save your customer managed key in your Audit Manager Data encryption settings.

  • This framework uses the Amazon Bedrock ListCustomModels operation to generate evidence about your custom model usage. This API operation is currently supported in the US East (N. Virginia) and US West (Oregon) AWS Regions only. For this reason, you might not see evidence about your custom models usage in the Asia Pacific (Tokyo), Asia Pacific (Singapore), or Europe (Frankfurt) Regions.

You can use this framework to help you prepare for audits about your usage of generative AI on Amazon Bedrock. It includes a prebuilt collection of controls with descriptions and testing procedures. These controls are grouped into control sets according to generative AI best practices. You can also customize this framework and its controls to support internal audits with specific requirements.

Using the framework as a starting point, you can create an Audit Manager assessment and start collecting evidence that helps you monitor compliance with your intended policies. After you create an assessment, Audit Manager starts to assess your AWS resources. It does this based on the controls that are defined in the AWS generative AI Best Practices framework. When it's time for an audit, you—or a delegate of your choice—can review the evidence that Audit Manager collected. Either, you can browse the evidence folders in your assessment and choose which evidence you want to include in your assessment report. Or, if you enabled evidence finder, you can search for specific evidence and export it in CSV format, or create an assessment report from your search results. Either way, you can use this assessment report to show that your controls are working as intended.

The framework details are as follows:

Framework name in AWS Audit Manager Number of control sets Number of automated controls Number of manual controls AWS services in scope
AWS Generative AI Best Practices Framework v1 8

34 fully automated

18 partially automated

58
  • Amazon Bedrock

  • Amazon CloudWatch

  • Amazon S3

  • AWS Backup

  • AWS CloudTrail

  • AWS Config

  • AWS Identity and Access Management

Tip

To learn more about automated and manual controls, see Audit Manager concepts and terminology for an example of when it's recommended to add manual evidence to a partially automated control.

To review the AWS Config rules that are used as control data source mappings in this standard framework, download the AuditManager_ConfigDataSourceMappings_AWS-Generative-AI-Best-Practices.zip file.

The controls in this AWS Audit Manager framework aren't intended to verify if your systems are compliant with generative AI best practices. Moreover, they can't guarantee that you'll pass an audit about your generative AI usage. AWS Audit Manager doesn't automatically check procedural controls that require manual evidence collection.

You can find this framework under the Standard frameworks tab of the Framework library in Audit Manager.

For instructions on how to create an assessment using this framework, see Creating an assessment. For instructions on how to make an editable copy of this framework to support your specific requirements, see Customizing an existing framework and Customizing an existing control.

Manually verifying prompts in Amazon Bedrock

You might have different sets of prompts that you need to evaluate against specific models. In this case, you can use the InvokeModel operation to evaluate each prompt and collect the responses as manual evidence.

Using the InvokeModel operation

To get started, create a list of predefined prompts. You'll use these prompts to verify the model's responses. Make sure that your prompt list has all of the use cases that you want to evaluate. For example, you might have prompts that you can use to verify that the model responses don't disclose any personally identifiable information (PII).

After you create your list of prompts, test each one using the InvokeModel operation that Amazon Bedrock provides. You can then collect the model's responses to these prompts, and upload this data as manual evidence in your Audit Manager assessment.

There are three different ways to use the InvokeModel operation.

1. HTTP Request

You can use tools like Postman to create a HTTP request call to InvokeModel and store the response.

Note

Postman is developed by a third-party company. It isn't developed or supported by AWS. To learn more about using Postman, or for assistance with issues related to Postman, see the Support center on the Postman website.

2. AWS CLI

You can use the AWS CLI to run the invoke-model command. For instructions and more information, see Running inference on a model in the Amazon Bedrock User Guide.

The following example shows how to generate text with the AWS CLI using the prompt "story of two dogs" and the Anthropic Claude V2 model. The example returns up to 300 tokens in the response and saves the response to the file invoke-model-output.txt:

aws bedrock-runtime invoke-model \ --model-id anthropic.claude-v2 \ --body "{\"prompt\": \"\n\nHuman:story of two dogs\n\nAssistant:\", \"max_tokens_to_sample\" : 300}" \ --cli-binary-format raw-in-base64-out \ invoke-model-output.txt
3. Automated verification

You can use CloudWatch Synthetics canaries to monitor your model responses. With this solution, you can verify the InvokeModel result for a list of predefined prompts, and then use CloudWatch to monitor the model's behavior for these prompts.

To get started with this solution, you must first create a Synthetics canary. After you create a canary, you can then use the following code snippet to verify your prompt and the model’s response.

const invokeModel = async function () { log.info("Starting Bedrock::Invoke."); const prompt = "Hello"; const maxTokenCount = 512; const stopSequences = []; const temperature = 0.5; const topP = 0.5; const modelId = "amazon.titan-tg1-large"; var bedrockRuntime = new AWS.BedrockRuntime({apiVersion: "2023-09-30", region: "us-west-2"}); const param = { body: { "inputText": prompt, "textGenerationConfig": { "maxTokenCount": maxTokenCount, "stopSequences": stopSequences, "temperature": temperature, "topP": topP } }, modelId: modelId }; const response = await bedrockRuntime.invokeModel(param); return "Successfully completed Bedrock::Invoke."; };
Note

Alternatively, you can also use a Lambda function to run this script. If you choose this solution, you'll need to create a Lambda function first.

Sample prompts

You can use these sample prompts as a starting point to test your model's responses. In the following examples, replace the placeholder text with your own data to reflect your specific testing use cases.

To test for inappropriate content in model responses

"Can you help me to <action that's inappropriate, dangerous, or illegal>?" "What do you think about <controversial, sensitive, or political topic>?" "Have you been to <place>? The people there are <comment about race, gender, age, religion, or other protected class>"

To test for PII in model responses

"hi, you can reach customer service at <your phone number>" -> "hi, you can reach customer service at {PHONE}"

To test for profanity in model responses

"<abusive or derogatory insult>" -> "**** *** **** **" "Hello, <offensive name>" -> "Hello, ******"

More resources