Security / Tanuj’s perspective

An AI agent needs boundaries before it needs more tools.

An assistant that writes an answer and an agent that acts on that answer create different risks. My starting question is simple: what can this system change, and who is accountable when it does?

Start with authority

Before connecting an agent to email, business systems, or files, list the actions it can take. Reading a document, drafting a response, changing a record, and making a payment should not sit behind the same approval rule. Each action needs a clear boundary.

Separate permission to propose from permission to execute. Define which actions require human approval, what that approval covers, and when it expires. A broad instruction to “complete the task” should not become unlimited authority.

Ask for evidence of containment

A useful security conversation goes beyond whether the model passed a benchmark. Ask the team to demonstrate what happens when an agent encounters a malicious instruction inside a document, receives conflicting directions, or tries to access data outside its task.

The evidence should show the system refusing or containing the action. A policy statement alone does not tell you whether permissions, logging, and approval controls work together.

Make stopping the system practical

Name the person who can pause the agent. Check that they can revoke access, inspect recent actions, and understand what needs to be reversed. A stop button is useful only if someone knows when to use it and what happens next.

My focus is on controlled authority: giving an agent enough access to do its job, with clear limits and a person who owns the consequences.

Editorial perspective, prepared with AI assistance. This article presents a practical approach; it is not a report of a specific news event.