Amazon has launched the InvokeGuardrailChecks API as part of its Amazon Bedrock Guardrails suite, enabling developers to apply individual safety checks at any point in agentic AI applications. The API allows for the invocation of supported safeguards without creating dedicated guardrail resources, offering flexibility in multi-turn workflows. Developers can define custom thresholds and actions, such as block, bypass, retry, or log results, based on specific requirements. This new API enhances the safety of generative AI applications by detecting and filtering undesirable content and protecting sensitive information in both user inputs and model responses.
The InvokeGuardrailChecks API operates in detect-only mode, returning numeric scores for each safeguard. These scores help developers implement context-aware logic in their application logic, such as blocking high-confidence threats or routing ambiguous findings to human review. The API uses a structured messages schema, with each content block having a required role such as system, user, or assistant, which provides the context needed for precise evaluation of content. This is particularly important for multi-turn agentic workflows where each step carries a different risk profile.
Amazon Bedrock Guardrails provides configurable safeguards to help developers build safe generative AI applications. With comprehensive safety controls across foundation models, the service helps detect and filter undesirable content and protect sensitive information in both user inputs and model responses. The new InvokeGuardrailChecks API extends these capabilities for agentic AI applications with multi-turn workflows, offering granular, per-request control over which safeguards to run at each stage of the agent loop.
Source: awsml
