Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarAWS Machine Learning Blog→ original

Skyflow in AgentCore Gateway hides personal data from AI agents' tools

When an AI agent searches the web for details on a bank customer's dispute, the customer's name goes into a query to a third-party search engine. Skyflow plugs into Amazon Bedrock AgentCore Gateway and replaces the name with a reversible token such as [NAME_AJbrIud]. The search engine sees the token, while the analyst sees the real name in the agent's reply. Neither the agents nor the tools need code changes, so it makes sense to set up protection once, at the gateway.

AI-processed from AWS Machine Learning Blog; edited by Hamidun News
Skyflow in AgentCore Gateway hides personal data from AI agents' tools
Source: AWS Machine Learning Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceAWS Machine Learning Blog ↗
Editor-in-Chief
Three modesprotect the tool, the model, or the document
Key metric #2If de-identification fails, the call is blocked rather than sent in the clear
Key metric #3PDFs and images get only irreversible tokens

Agents long ago stopped merely answering questions: they query search engines, CRMs, cloud storage and MCP servers. Every one of those calls is a channel through which personal data can leak. Most corporate safeguards watch what goes into and comes out of the model, but they rarely pay attention to what the agent sends to tools and receives back. AWS and Skyflow have proposed closing this gap in a single place: the Amazon Bedrock AgentCore Gateway.

Where exactly the data leaks

AWS describes three scenarios. First, the model builds the arguments of a call from its entire context, including material it does not need. Ask an agent about a customer, and the customer's name, account number or e-mail can end up in a request to a third-party server outside your account and outside your compliance boundary. Second, tool results, such as scraped pages or CRM records, settle into the model's context, memory, traces and logs, even when the task did not need them. Third, documents. Storage services such as S3, Box and Google Drive return not text but a signed link. A filter that inspects strings sees just a URL, lets it through, and the agent opens the original file.

The article's example is a bank agent investigating a disputed card transaction. Without protection, the search engine receives the query "Mateo Jackson dispute AnyCompany Electronics London", so the customer's name goes to a third party for a search that only needs the merchant. With the interceptor, the query reads "[NAME_AJbrIud] dispute AnyCompany Electronics London". The search results come back unchanged, the search engine does not know who the query is about, and the analyst still sees the real name in the agent's reply.

How the interceptor works

AgentCore Gateway lets you attach your own Lambda function to every call: a request interceptor sees the call before it reaches the tool, and a response interceptor sees the result before it reaches the agent. Either one can inspect, rewrite or block the content. Skyflow sits at both points and uses Skyflow Detect to find names, e-mails, phone numbers, government IDs, card numbers, dates of birth and addresses. The originals are stored in Skyflow Vault, so authorized systems can turn tokens back into real values.

There are three modes. "Protect the tool": sensitive values are removed from arguments and restored in results, which suits web search and third-party APIs. "Protect the model": prompts, history and results are de-identified before the model reads them; for this, the agent routes model calls through the gateway's inference endpoint. "Protect the document": the signed link is replaced with a link to a sanitized copy, which Skyflow Detect v2 creates and stores in an approved region. The file itself does not pass through Lambda, so payload size limits are not a problem.

The policy at the server and tool level (direction, entity types, token type, either reversible or one-time) lives in Parameter Store and can be changed without redeploying. The API key is kept in Secrets Manager, and decisions are logged to CloudWatch without the original values. If de-identification cannot be completed, the call is held back rather than sent in the clear. Files are fetched only from approved locations.

Limitations and what to do

There are caveats. Text and office formats (TXT, JSON, CSV, XML, JSONL, DOCX, XLSX, PPTX) keep reversible tokens, while PDFs and images are fully sanitized but their tokens are irreversible. String handling is generally available, but document de-identification on Skyflow Detect v2 is still in preview, so measure latency on your own data and region. Processing time grows with file size, so set a size limit that fits your real documents. Finally, this is a vendor solution: Skyflow is sold through AWS Marketplace, although the interceptor code is open source under the MIT license in the agentcore-skyflow repository.

The authors' practical advice is simple: pick one agent scenario that touches customer data and calls at least one external tool, list the data types that must not leave, and attach the interceptor to its gateway. Start with the tool mode, then add documents and the model, and watch the CloudWatch metrics for a week before expanding the rollout. For teams that already run several agents on different frameworks, the value is that the rule is set once instead of being baked into each agent separately.

Story

Key Questions Answered

Where exactly the data leaks?

AWS describes three scenarios. First, the model builds the arguments of a call from its entire context, including material it does not need. Ask an agent about a customer, and the customer's name, account number or e-mail can end up in a request to a third-party server outside your account and outside your compliance boundary. Second, tool results, such as scraped pages or CRM records, settle into the model's context, memory, traces and logs, even when the task did not need them. Third, documents.

How the interceptor works?

AgentCore Gateway lets you attach your own Lambda function to every call: a request interceptor sees the call before it reaches the tool, and a response interceptor sees the result before it reaches the agent. Either one can inspect, rewrite or block the content. Skyflow sits at both points and uses Skyflow Detect to find names, e-mails, phone numbers, government IDs, card numbers, dates of birth and addresses. The originals are stored in Skyflow Vault, so authorized systems can turn tokens back into real values.

What are the key details on Limitations and what to do?

There are caveats. Text and office formats (TXT, JSON, CSV, XML, JSONL, DOCX, XLSX, PPTX) keep reversible tokens, while PDFs and images are fully sanitized but their tokens are irreversible. String handling is generally available, but document de-identification on Skyflow Detect v2 is still in preview, so measure latency on your own data and region. Processing time grows with file size, so set a size limit that fits your real documents.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…