Skyflow Interceptor on Amazon Bedrock AgentCore Gateway: Three Modes to Protect Agent Tool Calls, Document Feature in Preview
An AWS blog explains how the Skyflow interceptor attaches to AgentCore Gateway to de-identify agent tool calls; string handling is generally available, while document handling (Detect v2) is still in preview.
News date: About 7 min read
Contents
On 2026-11-19, the AWS Machine Learning Blog published a post on adding Skyflow Detect to Amazon Bedrock AgentCore Gateway to protect AI agent tool calls. Here, an "agent" is an AI program that carries out tasks on its own and calls external tools such as web search, document stores or third-party APIs. "De-identification" means swapping personal data such as names and email addresses for placeholder tokens, so external tools cannot tell who the data belongs to. For enterprises whose agents look up customer data and call external tools, the aim is to reduce, in one place, the risk of personal data leaking to third-party tools or into model context. Everything below is as stated by the AWS blog.
What changed? The data gap in agent tool calls
The AWS article identifies three gaps. First, sensitive data leaks through tool arguments. Second, sensitive data returns to the model context along with tool results. Third, document tools return a signed URL (a file link carrying access permission) rather than text, so a filter that inspects only strings lets it through, and the agent then opens the original file.
The article explains that AgentCore Gateway is the gateway between the agent and all its tools. Its "interceptors" can run your own code, as an AWS Lambda function, on every call. A request interceptor inspects the call before it reaches the tool, and a response interceptor inspects the result before it reaches the agent; both can inspect, rewrite or block. The Skyflow interceptor attaches at both of these points.
The article gives the example of an agent at a large European bank investigating card disputes: once enabled, the search provider receives '[NAME_AJbrIud] dispute AnyCompany Electronics London', where [NAME_AJbrIud] is a reversible vault token, while the analyst still sees the real name in the agent's answer.
How do the three protection modes differ?
AWS explains that the Skyflow gateway configuration decides which mode applies to each tool. The interceptor calls Skyflow Detect to find sensitive entities such as names, email addresses, phone numbers, government IDs, card numbers, dates of birth and locations. The original values are stored in Skyflow Vault, a secure store through which authorized systems can restore tokens to the original values.
| Mode | How it works | Who can see real data | Typical use |
|---|---|---|---|
| Shield the tool | Sensitive values are removed from outbound arguments and restored in returned results; uses reversible vault tokens | The agent and the user, not the tool | Web search, data enrichment and third-party APIs that do not need customer identity |
| Shield the model | Prompts, history, files and tool results are de-identified before the model reads them, and re-identified when the answer comes back; model calls must go through the gateway inference endpoint | The agent and the user, not the model | Customer conversations, scraped web content, cross-region or third-party models |
| Shield the document | The signed URL is swapped for the URL of a de-identified copy; handled by Skyflow Detect v2 (preview), with files not passing through Lambda | No one on the agent side | Case notes, statements and contracts in Amazon S3, Box and Google Drive |
What are the release stage, supported formats and limits?
Per the article, string de-identification and re-identification are generally available. Document de-identification uses Skyflow Detect v2 and is still in preview. AgentCore Gateway availability varies by AWS Region, so check the supported Regions in the Amazon Bedrock AgentCore documentation before deploying. The article gives no information on availability in Taiwan.
| Item | What the article says |
|---|---|
| String handling | De-identification and re-identification are generally available |
| Document handling | Skyflow Detect v2, preview |
| Formats that keep reversible tokens | TXT, JSON, CSV, XML, JSONL, DOCX, XLSX, PPTX |
| Formats with one-way tokens only | PDF and images; original values cannot be restored from these files |
| File size | Checked before submission; files over the configured limit are rejected immediately rather than timing out |
How do you get started with deployment?
The AWS article lists the steps below. The interceptor is open source under the MIT license in the agentcore-skyflow repository on GitHub, and Skyflow is available on AWS Marketplace.
- Subscribe to Skyflow on AWS Marketplace, create a vault, and store the API key in AWS Secrets Manager.
- Use the AWS CDK to deploy the interceptor stack from the repository, which creates the Lambda function, its policy parameter, a DynamoDB table for model call records, and CloudWatch metrics and alarms.
- Attach the interceptor to the gateway at both interception points; if you manage the gateway with CloudFormation or CDK, declare it there so later updates preserve it.
- To protect the model, add an inference target and point the agent's model client at the gateway inference endpoint.
The article recommends picking one agent workflow that touches customer data and calls at least one tool outside your account, and listing the entity types that must never leave. Then start with Shield the tool and add the document and model modes afterwards. Before expanding, watch the CloudWatch metrics for a week. MCP servers outside the account (MCP is the common protocol agents use to connect to tools) can be set as gateway targets so their traffic also passes through the interceptor.
How does it fit with existing AgentCore controls?
AWS says Policy (using Cedar and Dogwood) decides what an agent may do, while Skyflow controls how sensitive data is handled when it does. Amazon Bedrock Guardrails handles content safety and sensitive information redaction, and can be combined with Skyflow for layered protection.
Frequently asked questions
Do I need to change the agent or tool code?
According to the AWS article, agents and tools need no changes; however, to protect the model, the agent must point its model client at the gateway inference endpoint.
Do policy changes require redeployment?
The article says policies are stored in AWS Systems Manager Parameter Store, so changes need no redeployment, and direction, entity types and token types can be set per server and tool.
How do reversible tokens differ from one-way tokens?
The article explains that vault tokens like [NAME_AJbrIud] can be restored by authorized users through Skyflow, while one-way counter tokens like [NAME_1] cannot be restored.
How is data residency handled?
The article says Skyflow offers several vault deployment models and Region choices to meet residency requirements; see Skyflow's deployment model documentation for details. The article gives no Taiwan Region information.
Browse the latest news in this topic
Found this guide helpful? Share it with someone who needs it
Claude Code 2.1.329: Code Mode Works With Telemetry, /tasks Lists Scheduled Tasks, macOS Computer Use Gets Stricter Safety Rules
Anthropic's changelog lists Claude Code 2.1.329, covering Code Mode behavior changes, new diagnostics in /tasks and /mcp, /code-review changes, and new safety rules for macOS computer use.
Kimi Code CLI v2.7.0 Adds a Native Amazon Bedrock Provider: Use Your AWS Credentials Instead of a Separate API Key
AWS Rounds Up 13 SageMaker Inference Features: Managed NVIDIA Dynamo on HyperPod, MIG GPU Partitioning, Cross-Region Routing
BMW Group Shares How It Governs AI Agents: 12,500+ Developers, Three Autonomy Levels and a Four-Stage CI/CD Gate
Sources
- Protect agent tool calls with Skyflow and Amazon Bedrock AgentCore Gateway | Artificial IntelligenceChecked:
Compiled from the references and verification dates listed above; fares, schedules, product features, and terms can change, so please confirm the latest details on the relevant official website before travelling, booking, or purchasing.