A company provides a GenAI application that uses Amazon Bedrock to customers.

Amazon Web Services AIP-C01 Question Answer

A company provides a GenAI application that uses Amazon Bedrock to customers. The application accepts untrusted user inputs. The company observes that some users attempt to bypass system instructions by using prompt injection and jailbreak techniques. The company needs a solution to protect the application from malicious actors.

The solution must meet the following requirements:

• Detect and mitigate adversarial user inputs before the application invokes the model.

• Enforce consistent safety controls during model inference.

• Prevent the application from returning unsafe or manipulated outputs to users.

• Use managed AWS services where possible to minimize the need for custom security logic.

Which solution will meet these requirements?

Use AWS Lambda functions to sanitize user inputs. Apply Amazon Bedrock Guardrails during model inference. Validate outputs before returning responses.

Use carefully engineered system prompts to discourage prompt injection and jailbreak attempts during model interactions.

Restrict access to the application by using IAM authentication. Assume authenticated users will not submit adversarial inputs.

Log all prompts and responses to Amazon S3. Query the prompts and responses periodically to identify adversarial behavior.

Explanation:

Option A is the only choice that establishes controls across the complete input-inference-output path. Amazon Bedrock Guardrails provides managed safeguards that evaluate model prompts and responses. AWS explicitly supports a PROMPT_ATTACK content-filter category designed to detect jailbreaks, prompt injection, and—with the appropriate tier—prompt leakage. Prompt injection attempts try to override developer instructions, while jailbreaks attempt to circumvent native model safety controls.

When using Guardrails with model invocation APIs, applications can mark untrusted user content with guardrail input tags so that the prompt-attack filter evaluates the user-controlled portion without incorrectly treating developer or system instructions as adversarial content. Guardrails can then block detected prompt attacks before unsafe content is processed. Content filters and other guardrail policies can also inspect generated model responses before they reach the end user.

AWS also recommends associating a Guardrail with Amazon Bedrock agents to help protect against prompt injection. System prompts remain useful as a defense-in-depth technique because they define the agent ' s scope, but AWS does not treat system prompting alone as an adequate security boundary.

The Lambda sanitization and output-validation portions of A can provide additional application-specific checks. In a modern implementation, Guardrails itself can perform substantially more of the input and output filtering, reducing the amount of custom Lambda logic required.

B relies entirely on probabilistic prompting and therefore lacks deterministic managed enforcement. C is fundamentally unsafe because authenticated users can still be malicious or compromised. D is reactive monitoring: logging and periodically searching historical prompts may help investigations, but it does not prevent an attack before inference or block unsafe responses.

Therefore, A is the only listed design that combines pre-processing controls, managed inference-time safeguards, and output validation.

==========

Amazon Web Services AIP-C01 View All Questions