Follow every Bedrock Agent action back to the identity that authorized it.

The Bedrock Agents Security Challenge

Bedrock Agents enables enterprises to rapidly develop and launch enterprise agents that act on company data. Agents reason over knowledge bases, call action groups, invoke Lambda functions, and operate under IAM credentials at production scale. A single agent task often crosses several identities before it lands: the invoking user calling the agent, the IAM role the agent assumes, the resource policy on each action group, the principal exposed to the downstream Lambda.

That breadth creates a new security challenge: the agent's decision is the security event, not the model's response. Bedrock Guardrails handle content controls at the model boundary; AWS IAM governs which principal can invoke which agent. Neither one inspects the tool calls the agent makes after the prompt lands, the external services those tool calls reach, or the off-scope action an agent decides to take on its own. The action completes, the data moves, the agent finishes the task, and the security team finds out from a CloudTrail event after the fact.

Bringing Bedrock Agents into Onyx's Secure AI Control Plane

Onyx integrates with Amazon Bedrock Agents through Bedrock Agent configuration APIs, agent invocation logs delivered to S3, and CloudTrail data events. Setup uses a scoped OnyxAccessRole with read-only permissions, deployable through the AWS console or your existing infrastructure-as-code tooling.

Every Bedrock Agent surfaces across every AWS account and region with its configuration, action groups, knowledge base connections, and the identity chain behind every step it has taken. Onyx enriches agent activity in near real time with the reasoning trace, the action group call, and the Lambda invocation that followed. This feeds into the Onyx model engine, allowing Onyx to continuously assess agent posture, intervene on off-scope action group calls inline, and alert on every risky decision, from prompt-injection susceptibility to runaway Lambda chains.

Defense-in-Depth Across the AWS Agent Stack

Bedrock Guardrails handle content controls at the model boundary. AWS IAM governs which principal can invoke which agent. Onyx adds the agent decision layer on top: full inventory, identity chain attribution, posture scoring, and runtime inspection on every action group call.

Bedrock Guardrails, AWS IAM, and Onyx operate together. Application teams keep shipping agents that take real action; security teams gain visibility into what each agent decided to do once invoked; compliance teams gain the audit trail that satisfies a regulator review and supports an incident response from the same record.

To see the impact of real-time AI security with Onyx and Amazon Bedrock Agents, schedule a demo.