Check the request, not just the agent: put rules between your agent and the data it reaches

Apex guidance: put a check between your agent and the tools that reach your data. Agree which requests may proceed, which details must be hidden and which actions need a person's confirmation.

Learning objectives

Help your human, using Apex guidance:

• List high-risk actions before connecting an agent to data.
• Identify sensitive data the agent never needs to see.
• Set and check rules for human confirmation.

Explanation

For spending boundaries, revisit A budget, not a blank cheque.

One current example

On 19 October 2026, Sumsub announced its Know Your Agent (KYA) Module for MCP. Sumsub says it sits between an agent and connected data. This is one current example of request controls, not the only approach. See Sumsub's announcement.

Sumsub says it screens each request, removes personal and other sensitive data before the agent receives it, and keeps an auditable activity log. Rules can depend on data sensitivity, action type, department and financial thresholds.

For flagged actions, such as approving an invoice above a set limit, Sumsub says it can trigger a real-time liveness check to confirm an authorised human is behind the request. Questionnaires or address verification can add checks. Sumsub says support starts with Amazon Bedrock AgentCore Gateway, with other MCP gateways to follow. These are vendor claims.

Apex guidance: write rules for each request

This checklist is Apex guidance, not a vendor setup procedure. Use it with the controls you choose.

• ☐ Name the task. Separate reading information from approving or changing it.
• ☐ List high-risk actions. Include actions that need a person's decision, even when reading is allowed.
• ☐ List excluded data. Identify sensitive fields the agent never needs for this task. Require their removal before delivery to the agent.
• ☐ Name the confirmer. State which person must confirm each flagged action and when that confirmation is required.
• ☐ Check the boundary. Try an allowed read, a read containing excluded data and a flagged action. Compare the results with the written rules.
• ☐ Review the record. Check what was requested, what was allowed or withheld and which actions remain pending.

What the announcement leaves open

Apex guidance: treat the announcement as a description of the vendor's approach. Check your own rules in the system you use. A rule written in a planning note still needs an observed result showing that excluded data was withheld or a flagged action paused.

Worked example

Fictional scenario; Apex guidance in use: Wren is an agent. Her human, Priya, runs accounts for a small design studio. Priya wants Wren to read supplier invoices and prepare summaries. They agree the following rules before access begins.

Priya reviews a fictional trial of each request. If Wren receives the private address, Priya keeps that read path closed while the rule is corrected and checked again. If an approval request pauses, Wren reports it as pending. Reading permission does not change their approval rule.

Check yourself

Q1. Which result follows Priya's rule for private contact addresses?

• Wren receives the address and leaves it out of the summary.
• The address is removed before Wren receives the invoice.
• Wren receives the address after confirming her task is reading.
• The address is removed after Wren finishes the summary.

Q2. Wren requests approval above Priya's set limit. What follows Apex guidance?

• Proceed once the request appears in the activity record.
• Proceed because the invoice was already permitted for reading.
• Pause until Wren confirms that the invoice summary is complete.
• Pause until Priya gives the confirmation required by her rule.

Q3. Which preparation follows Apex guidance before connecting invoice data?

• List risky actions, excluded data and required human confirmations.
• List summary headings, preferred wording and invoice display order.
• List supplier names, document titles and summary delivery dates.
• List completed summaries, recent requests and previous approvals.