AWS and Skyflow add data protection to AI agent tool calls

AWS and Skyflow have demonstrated a way to keep sensitive information out of the tools used by AI agents and out of the model context those tools provide. The approach adds Skyflow Detect as a request and response interceptor to Amazon Bedrock AgentCore Gateway.
- Skyflow Detect can de-identify request and response data at AgentCore Gateway.
- The integration supports Shield the tool, Shield the model and Shield the document modes.
- Agents and tools do not need code changes for the interceptor setup.
- The interceptor can fail closed when data protection cannot be completed.
- Skyflow is available in AWS Marketplace and the interceptor is open source on GitHub.
Why agent tool calls create a data gap
AI agents increasingly act through Model Context Protocol servers, SaaS APIs, document stores and web search. AWS says every tool call is a data flow, but many enterprise controls focus on what enters and leaves the model rather than what an agent sends to a tool or receives back.
That gap can expose sensitive information in several ways. An agent may include a customer’s name, account number or email address in a query sent to a third-party server. Tool results such as CRM records, case notes and scraped web pages can then enter the model’s context, memory, traces and logs.
Documents create another challenge. Tools for Amazon S3, Box and Google Drive can return signed URLs instead of text. A filter that inspects text may allow the link through, after which the agent opens the original file.
Interceptors sit between agents and their tools
Amazon Bedrock AgentCore Gateway fronts the tools used by an agent, regardless of the framework or model behind it. Its interceptors can run code on every call passing through the gateway. A request interceptor sees a call before it reaches a tool, while a response interceptor sees the result before it reaches the agent. Either can inspect, rewrite or block the payload.
The Skyflow integration uses one AWS Lambda function at both interception points. Skyflow Detect identifies entities including names, email addresses, phone numbers, government IDs, card numbers, dates of birth and locations. It can replace those values with tokens or restore the originals. Skyflow Vault stores the original values so authorized systems can reverse reversible tokens.
AWS says the setup does not require changes to agent or tool code. The interceptor can apply one control across every agent and tool connected to the gateway, alongside policy and guardrail controls.
Three protection modes for different data flows
The gateway configuration supports three protection modes. Shield the tool removes sensitive values from outgoing arguments and restores them in returned results. This is intended for web search, enrichment and third-party APIs that do not need customer identity.
Shield the model de-identifies prompts, history, files and tool results before the model reads them, then re-identifies the answer on the way back. This mode can be used for customer conversations, scraped web content, and calls to cross-Region or third-party models.
Shield the document replaces a file’s signed URL with a URL for a de-identified copy. Skyflow Detect v2, identified in the source as a preview, fetches the file, cleans it and stores the copy in an approved Region. The file does not pass through Lambda, so payload limits do not cap its size.
A bank dispute workflow illustrates the approach
AWS describes a hypothetical dispute-resolution agent at a large European bank. An analyst asks it to investigate a disputed transaction. The agent searches for the merchant and retrieves a case note from Amazon S3.
Without protection, the search provider could receive a query containing the customer’s name, such as ‘Mateo Jackson dispute AnyCompany Electronics London’. The agent could also open the original case note, including the customer’s full name, email, phone number and card details.
With the Skyflow interceptor, the name is replaced in the search query by a reversible vault token, shown in the source as ‘[NAME_AJbrIud] dispute AnyCompany Electronics London’. The provider does not learn the customer’s identity, while the analyst can still see the real name in the agent’s answer. Web content reaches the model de-identified, and the agent receives a link to a de-identified copy of the case note.
The design also supports policies stored in AWS Systems Manager Parameter Store, with changes taking effect without redeployment. AWS says the interceptor fails closed when de-identification cannot be completed, fetches files only from allowlisted locations, stores the Skyflow API key in AWS Secrets Manager, and logs decisions to Amazon CloudWatch without logging credentials or original values. Skyflow is available in AWS Marketplace, and the interceptor is open source on GitHub.


