Protect agent tool calls with Skyflow and Amazon Bedrock AgentCore Gateway
Learn how Skyflow and Amazon Bedrock AgentCore Gateway keep sensitive data out of the tools your AI agents call. Added as a gateway interceptor, Skyflow de-identifies data before any tool, model, or log sees it, and restores it for authorized callers, with no change to your agent code.
In this post, we explore how Skyflow and Amazon Bedrock AgentCore Gateway safeguard sensitive data by intercepting tool calls made by AI agents. Skyflow Detect is integrated as a request and response interceptor within the gateway. The interceptor de-identifies data prior to it reaching a tool, restores it afterward, and substitutes documents with clean copies before the agent accesses them.
Agents and tools remain unchanged. Skyflow is accessible via AWS Marketplace, and the interceptor is open source on GitHub. Agents communicate via tools, and tools manage data as agents transition from answering inquiries to executing actions. These interactions involve tool arguments, tool results, and documents. Most enterprise controls focus on data entering and exiting the model, but fewer address data exchanged between agents and their tools, and the data those tools return.
This oversight manifests in three ways: sensitive data leaving via tool arguments, data entering the model through tool results, and documents bypassing text inspection when converted into signed URLs. These issues can't be mitigated by patching individual agents or tools, as they span teams and frameworks. A common solution is to place controls at every tool call, ensuring consistent authorization, content safety, and data protection.
AgentCore Gateway serves as this central point, fronting every tool an agent utilizes, regardless of the framework or model in use. With AgentCore Gateway interceptors, custom code can be executed as an AWS Lambda function for each call passing through the gateway. A request interceptor examines each call before reaching a tool, while a response interceptor examines each result before it reaches the agent.
Through these interceptors, data protection can be integrated alongside Policy and Guardrails. This unified control applies to all agents and tools connected to the gateway without requiring modifications to the agents or tools. For instance, consider a European bank employing agents to investigate card disputes. During a dispute investigation, an analyst queries the agent to research a disputed transaction.
The agent performs a web search for the merchant, retrieves the dispute case note from Amazon S3, and summarizes it. Without protection, the following occur: the model generates a search query from the agent's context, transmitting the customer's name to a third-party search provider; the search provider receives personal data unrelated to the search; and the agent accesses the full case note, including the customer's name, email, phone number, and card details.
With the Skyflow interceptor in place on the gateway, the workflow changes. The search provider receives a reversed token in place of the customer's name, preventing the provider from learning the customer's identity. The merchant results remain unchanged, and the analyst continues to see the customer's name in the agent's response.
Instead of the original case note, the agent receives a de-identified version, protecting the customer's sensitive information.
Written by urgent.news from AWS Machine Learning's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.