Protect agent tool calls with Skyflow and Amazon Bedrock AgentCore Gateway
Skyflow and AWS integrated Skyflow Detect with Bedrock AgentCore Gateway to redact sensitive data from autonomous agent tool calls using reversible tokens.
By Dillip Chowdary • Nov 20, 2026 • Source: AWS Machine Learning Blog

At a glance
- Product
- Skyflow with AgentCore Gateway
- Document engine
- Skyflow Detect v2
- Availability
- AWS Marketplace and GitHub
- Policy store
- AWS Systems Manager Parameter Store
- Key storage
- AWS Secrets Manager
- Audit logging
- Amazon CloudWatch
In this post
Amazon Web Services and Skyflow introduced an integration connecting Skyflow Detect to Amazon Bedrock AgentCore Gateway to secure artificial intelligence agent tool calls and model context. The architecture deploys an open source interceptor that strips sensitive personal records before outbound tool requests leave an enterprise perimeter, sanitizes incoming tool results before an agent consumes them, and replaces private document links with de-identified counterparts. Enterprise teams can source the Skyflow service through AWS Marketplace while deploying the gateway interceptor code directly from GitHub, as detailed in AWS Machine Learning Blog's report.
This report examines how the interceptor prevents data exfiltration during autonomous agent operations across third-party software interfaces, cloud storage, and large language models. The technical walkthrough applies to cloud architects, data security officers, and enterprise developers building autonomous agents that query external systems under stringent data residency and privacy mandates. Operations teams can enforce corporate compliance rules across distributed agent frameworks without modifying underlying model logic or individual tool implementations.
01Enterprise data leakage across agent tools
Autonomous agents execute complex tasks through external utilities, including Model Context Protocol servers, software-as-a-service application programming interfaces, cloud document stores, and web search engines. While standard enterprise safeguards routinely filter data entering and leaving the primary foundation model, conventional perimeters fail to regulate what agents pass directly to their execution tools. Every automated tool invocation represents an unmonitored data pathway that can bypass internal compliance controls.
Sensitive customer records frequently leave corporate accounts because generative models construct tool parameters using their entire conversational history rather than minimal query inputs. When an agent queries an external database about an account holder, extraneous details like national identity numbers, residential addresses, and personal phone numbers get bundled into outbound payload strings. These uninspected queries land on third-party servers operating outside local compliance boundaries and cloud tenancy controls.
Outbound data leakage compounds when third-party tools transmit unvetted payloads back into the system. Scraped website data, enterprise customer relationship management logs, and internal case summaries enter conversational context, persistent agent memory, operational traces, and system diagnostics. Furthermore, document tools hosted on Amazon Simple Storage Service, Box, and Google Drive return signed uniform resource locators instead of raw text, allowing uninspected files to bypass conventional string inspection filters.
02Request and response interceptor architecture
Amazon Bedrock AgentCore Gateway serves as a unified entry point fronting every tool an agent invokes, regardless of the underlying orchestration framework. By incorporating request and response interceptors built as AWS Lambda functions, the gateway inspects and modifies payloads passing in either direction. The single proxy configuration applies uniform content safety, authorization rules, and data residency policies across disparate tool inventories without requiring custom application code updates.
The deployed Lambda function coordinates with Skyflow Detect to identify sensitive entities across multiple data classifications, including personal names, email addresses, government identifiers, payment card numbers, birth dates, and geographical locations. Sensitive values are substituted with deterministic tokens before downstream processing, while Skyflow Vault stores the original records within an approved compliance region. Authorized endpoints can later exchange these tokens to recover the source values before displaying completed summaries to human operators.
System administrators configure entity handling rules, traffic direction, and token formats through AWS Systems Manager Parameter Store without triggering service redeployments. Skyflow authentication credentials reside securely within AWS Secrets Manager, while all operational routing choices generate audit logs inside Amazon CloudWatch. The interceptor operates on a fail-closed principle, withholding tool transmissions entirely if data redaction fails or if unapproved file repositories attempt to bypass inspection checks.
03Disputed card transactions at European banks
Advertisement
Tech Pulse Daily
Get tomorrow's pulse first
Join engineers who read Tech Pulse before stand-up. Free, weekday mornings.
A multinational European financial institution processing retail and commercial banking accounts illustrates the operational risk during payment card dispute investigations. Compliance mandates obligate the bank to preserve strict data residency boundaries while internal analysts employ autonomous agents to assess disputed transactions. When an analyst asks an agent to investigate a merchant claim, the agent initiates web searches for business credentials and retrieves case documentation from Amazon Simple Storage Service.
An unmonitored agent pulls the cardholder name Mateo Jackson into the external search query along with the commercial vendor AnyCompany Electronics London. Without protective screening, the third-party search provider captures the customer identity alongside the merchant inquiry, while scraped search listings introduce unvetted third-party text into model context. The agent subsequently reads the raw storage record containing unredacted payment card sequences, billing addresses, telephone numbers, and email accounts.
Deploying the gateway interceptor replaces the cardholder identity with an isolated reversible token formatted as [NAME_AJbrIud] before dispatching the outbound search payload. The external engine executes the merchant lookup without learning who initiated the transaction, and web search responses undergo de-identification before entering memory. The analyst receives a final summary displaying the original customer name because authorized gateway routines restore vaulted values upon presentation.
04Three operational gateway protection modes
The integration organizes data protection into three distinct operational modes based on tool sensitivity and data consumption patterns.
| Mode | What happens | Who sees real data | How it works | Typical use |
|---|---|---|---|---|
| Shield the tool | Sensitive values are stripped from arguments on the way out and restored in results on the way back | The agent and user, not the tool | Uses reversible vault tokens, so values the tool echoes back are restored before the agent reads them | Web search, enrichment, and third-party APIs that don’t need customer identity to do their job |
| Shield the model | Prompts, history, files and tool results are de-identified before the model reads them. The answer is re-identified on the way back | Agent and user, not the model | Model calls go through the gateway’s inference endpoint. On the way back, only the tokens minted for that call are restored | Customer conversations, scraped web content, cross-Region or third-party models |
| Shield the document | The file’s signed URL is swapped for one pointing to a de-identified copy | No one on the agent side | Skyflow Detect v2 fetches the file, cleans it and stores the copy in an approved Region. The file never passes through Lambda, so payload limits don’t cap file size | Case notes, statements, contracts in Amazon S3, Box, Google Drive |
Note: Skyflow Detect v2 operates in preview to sanitize documents fetched through presigned links without streaming heavy files across Lambda memory limits.
05Parameter Store policy and runtime flow
Agent interactions follow an end-to-end execution sequence coordinated by AgentCore runtime and AWS Identity and Access Management policies.
- IAM signed request The agent calls a tool through AgentCore Gateway, signed with its IAM role.
- Request interception The REQUEST interceptor matches the tool against policy and de-identifies arguments with Skyflow Detect.
- Tokenized execution The gateway calls the tool, which receives tokens instead of sensitive values.
- Response interception The RESPONSE interceptor re-identifies the result, de-identifies it, or swaps a document URL as required.
- Result delivery The agent receives the final processed result.
Model-bound workflows route prompts through the gateway inference endpoint, where the interceptor logs token allocations into Amazon DynamoDB. This short-lived tracking table records transient token pairs without storing raw text, enabling accurate re-identification when completion responses return to the caller. The architecture isolates external Model Context Protocol servers, Amazon Bedrock foundation models, and Amazon Simple Storage Service data buckets behind cryptographic token exchanges.
Frequently asked questions
Where does the Skyflow interceptor run within AWS infrastructure?
The interceptor runs as an AWS Lambda function attached directly to the request and response interception points of Amazon Bedrock AgentCore Gateway.
How does the system handle sensitive documents stored in Amazon S3 or Box?
Skyflow Detect v2 retrieves the file directly through a presigned URL, creates a sanitized copy within an approved region, and swaps the original link for the clean copy.
What happens if Skyflow Detect fails to sanitize an outbound tool call?
The gateway interceptor fails closed, blocking the outbound tool transmission entirely instead of allowing raw, unredacted records to leave the enterprise network.
Where are the original sensitive values stored when tokens are created?
Original sensitive values are stored inside Skyflow Vault in an approved compliance region, allowing authorized systems to reverse tokens when returning results to users.
Sources
Author
Dillip Chowdary
Writes Tech Bytes coverage of AI, engineering, and the tools that actually ship. Editor of Tech Pulse Daily.
Related on Tech Bytes
Advertisement