Skip to content

Shared responsibility for generative AI: what you still own

AWS secures the cloud; you own your data, access and AI behaviour. See the customer-side checklist for generative AI on Amazon Bedrock.

5 min readBy AI Guru AcademyFacts last checked

By AI Guru Academy

In short

AWS is responsible for security of the cloud, which means the infrastructure that runs its services. You are responsible for security in the cloud: your data, who can reach it, how you configure each service and the rules that apply to your business. Using a managed generative AI service moves some work to AWS, but it does not move your accountability for how the application behaves.

What does the AWS shared responsibility model say?

The model splits duties between AWS and the customer. AWS describes the split as security "of" the cloud versus security "in" the cloud.

AWS is responsible for protecting the infrastructure that runs the services offered in the AWS Cloud. The Amazon Bedrock security page adds that AWS also provides services you can use securely, and that third-party auditors regularly test and verify the effectiveness of AWS security as part of the AWS Compliance Programs.

The customer side is stated just as plainly. Your responsibility is determined by the AWS service you use. You are also responsible for the sensitivity of your data, your company's requirements and applicable laws and regulations. The AWS model page lists data, encryption, identity and access management (IAM) and configuration on the customer side. The exact split depends on the services used, so check the documentation for each one.

Does the model change for generative AI?

The principle is the same, but the things you must decide are different. A generative AI application adds prompts, model outputs, source documents and sometimes customised models. Each of these is your data or your configuration, so each sits on your side of the line.

The Amazon Bedrock documentation says it is written to help you apply the shared responsibility model when using that service. Its security section covers configuring Bedrock to meet your security and compliance objectives, detecting and filtering harmful content with Guardrails, and observability.

What exactly do customers still own?

You own four areas, whichever model you choose.

What exactly do customers still own?
AreaWhat you decideIllustrative question
DataWhat goes into prompts, knowledge sources and customisation jobsShould customer records be in this prompt at all?
PermissionsWho and what may call a modelWhich roles need access to which models?
Application behaviourWhat the assistant may say and how it is checkedWhat should it refuse to discuss?
ComplianceWhich laws, policies and evidence applyWhich regulations cover our users?

Encryption is a good example of a configuration choice. For model customisation jobs and their output custom models, Bedrock documents that you can use an AWS KMS customer-managed key. By default, Bedrock encrypts custom models with AWS owned keys, so choosing your own key is a decision you make and record.

Why is access control separate from content filtering and grounding?

They answer different questions, so one cannot stand in for another. Access control decides who may use a model. Content filtering decides what goes in and out. Grounding checks whether an answer is supported by source material.

Access control uses IAM. The Bedrock identity-based policy examples allow or deny invocation actions such as bedrock:InvokeModel on specific model ARNs to control which principals can invoke a model. To stop a user invoking foundation models, you deny the API actions that invoke models directly.

Content filtering is a Guardrails feature. Content filters detect and filter harmful text or image content in input prompts or model responses. The documented harmful categories are Hate, Insults, Sexual, Violence, Misconduct and Prompt Attack. Denied topics help block topics you define if they are detected in user queries or model responses. Sensitive information filters can help block or mask personally identifiable information (PII) in inputs and responses.

Grounding is a further check. The Guardrails grounding check is documented as detecting and filtering hallucinations in model responses when a reference source and a user query are provided. It evaluates responses against the source you supply, and the grounding threshold is configurable.

Notice what each control does not do. A permitted user can still ask a question that produces an unsupported answer. A grounded answer can still come from someone who should not have had access. You need all three layers, and each one is your configuration.

Worked example: a claims assistant (illustration)

Imagine a fictional insurer, Northfield Mutual, that builds an assistant to summarise claim notes for its own staff. This is an invented illustration, not a real company.

AWS secures the infrastructure under the service. Northfield Mutual then makes its own decisions:

  1. Only the claims team role may invoke the model, using IAM permissions scoped to that model.
  2. A guardrail masks personal information in responses and blocks a denied topic such as legal advice.
  3. A grounding check compares summaries with the claim notes passed in as the reference source.
  4. A customer-managed KMS key protects any customised model artefacts.
  5. A named owner records which regulations apply and who reviews the setup.

If a summary exposes a customer's details to the wrong person, AWS operating the infrastructure securely does not answer for it. Permissions and configuration were Northfield Mutual's to set.

One limit applies. The documentation states that the sensitive information filter does not evaluate tool call arguments or tool results in tool-use workloads, so PII there is neither blocked nor masked. It also states that masking does not apply to model invocation logs: if invocation logging is enabled, the logged input contains the original request. A customer who relies on masking must plan for those places separately.

What is the customer-side checklist?

Use this list before any generative AI feature goes live.

  • Classify the data that will appear in prompts, sources and customisation jobs.
  • Grant model access through least-privilege IAM, and deny it where it is not needed.
  • Decide which encryption keys protect customised model artefacts.
  • Configure content filters, denied topics and sensitive information handling for inputs and outputs.
  • Add a grounding check where answers must follow a reference source.
  • Identify where masking does not reach, such as logs and tool-use data.
  • Record applicable laws, internal policies and a named owner.
  • Review the AWS documentation for each service you use, because the split depends on the service.

Key takeaways

  • AWS secures the cloud; you secure what you put and run in it.
  • Your share is set by the service you use, so read each service's documentation.
  • You still own data, permissions, application behaviour and compliance.
  • IAM access control, Guardrails content filtering and grounding checks are separate controls for separate risks.
  • Masking has documented limits, so plan for logs and tool-use data.

FAQ

Who is responsible for securing a generative AI application on AWS?

Both parties, for different parts. AWS protects the infrastructure that runs its services, and you are responsible for your data, access and configuration, plus the laws and requirements that apply to you.

Does Amazon Bedrock Guardrails replace IAM permissions?

No. IAM controls which principals can invoke a model, while Guardrails filters content in inputs and responses. They address different risks and you configure both.

Is a grounding check the same as a content filter?

No. A content filter screens for harmful content categories. A grounding check evaluates a response against the reference source you provide.

How should I prepare for questions on this topic?

Learn who owns each control and why the controls are separate. Concept-level understanding of the split is more useful than memorising settings.

Sources

Last checked 2026-11-16.

AI Guru Academy is an independent training provider and is not affiliated with or endorsed by Amazon Web Services. AWS service names are trademarks of Amazon.com, Inc. or its affiliates.

Next step: take the free AIB-C01 diagnostic at /diagnostic, then see the full course at /courses/aib-c01-exam-prep.

Shared responsibility for generative AI: what you still own | AI Guru Academy