Amazon Bedrock Guardrails получил новый API для гибкой защиты агентных AI-приложений

Amazon Web Services has launched a new API for the Amazon Bedrock Guardrails platform — InvokeGuardrailChecks — which allows embedding individual safety checks into agentic AI applications at any point in the workflow and without needing to create guardrail resources in advance.

What the new API changed

Until now, using Amazon Bedrock Guardrails required upfront configuration: a developer would create a guardrail resource in the console, bind it to a model or endpoint — and only then could they use the protection mechanisms. This meant a centralized approach: one guardrail for the entire application or a large block.

InvokeGuardrailChecks changes the logic. Now you can call a specific safety check directly from code at any arbitrary point in the pipeline. A check for personal data leakage — at one point, filtering unwanted content — at another, verification of authenticity — at a third. Each call is independent and does not require an existing guardrail resource.

Why this matters for agentic systems

Modern agentic AI systems work fundamentally differently than a simple request-response flow. A single user request can trigger a chain of dozens of steps: searching a knowledge base, calling external APIs, generating intermediate conclusions, making decisions based on the data received. At each of these nodes, problematic content or undesired behavior may appear.

Traditional guardrails, configured only at the input and output of the system, simply do not see what happens inside the multi-step process. InvokeGuardrailChecks closes this gap: the developer themselves chooses where a check is needed and embeds it as a regular API call.

The API supports the following types of checks:

• Detection and masking of PII (personal data)
• Content filtering by customizable categories
• Blocking forbidden topics (topic denial)
• Verification of response credibility based on sources (grounding)
• Detection of prompt injection attempts

Application in enterprise scenarios

The new API is especially relevant for enterprise applications where an agent works with sensitive data or has access to internal corporate systems. In HR bots, it is important to ensure that the agent does not disclose one employee's personal data to another — even if it appeared in an intermediate LLM output. InvokeGuardrailChecks allows adding a PII check at exactly the moment when the agent is assembling its response.

In financial applications, a grounding check is embedded before the final conclusion — so the agent relies on real sources rather than generating unreliable data. In multi-agent systems, a prompt injection check is added at the stage of receiving data from an external agent.

"With this API, developers can apply individual security checks at any point in agentic AI applications — without needing to create guardrail resources," — AWS

Machine Learning Blog.

What this means

Amazon is consistently building enterprise infrastructure for the safe deployment of AI agents. InvokeGuardrailChecks is a logical continuation of this strategy: instead of one centralized barrier, the company offers a set of targeted tools that developers embed where control is truly needed. For industries with strict compliance requirements — finance, healthcare, law — this is a serious step toward practical application of agentic AI without losing control over system behavior.