Safeguard your agentic AI applications with the Amazon Bedrock Guardrails InvokeGuardrailChecks API
TL;DR
AWS announced InvokeGuardrailChecks for Amazon Bedrock Guardrails on June 16, 2026. Developers can run individual safety checks inside agent loops without creating guardrail resources or managing versions. The API is detect-only. It does not block, mask, or rewrite content, but returns discrete scores from 0 to 1. The application then decides whether to allow, block, retry, escalate, or log.
Nauti's Take
The packaging is PR-heavy, but the direction is useful. Agent safety often fails when teams place one large guardrail block before or after the model and hope every intermediate step stays clean.
Detect-only is the more honest design here: AWS provides signals, while the actual policy stays inside the product. That is where it belongs, because a banking agent should not react like a creative writing tool.
Briefingshow
Agents do not have one safety problem. User input, system prompts, tool output, and final responses often need different checks. InvokeGuardrailChecks makes those controls more granular, but it also moves more responsibility to the product team: thresholds, escalation paths, and logging need real design work.
Sources
