Give Your AI Agent a Job Description, Not a Master Key: How to Control What It Can Access

By Dave Millier / November 18, 2026

Treat an AI agent like a new hire

Imagine hiring someone on Monday and handing them the keys to every system, every inbox and the company bank account before they have met the team. Nobody would do that. Yet giving an AI agent wide access is the easiest way to get it working, and it is one of the first things government security agencies tell organizations to avoid. An AI agent is a tool that does more than answer questions. It takes actions on its own, like approving a purchase or replying to a customer.

Canada's Cyber Centre, together with partner agencies in the United States, Australia, New Zealand and the United Kingdom, has published joint guidance on adopting these tools. Its advice on access is direct: never grant an agent broad or unrestricted access, especially to sensitive data or critical systems (Source: Canadian Centre for Cyber Security, Careful adoption of agentic AI). This post shows a simple way to follow that advice: write each agent a job description before you switch it on. For the wider picture on AI governance, see our earlier post, Get your AI house in order before it becomes someone else's problem.

Why access is the part to get right

An agent can only cause the harm its access allows. The joint guidance says the privileges given to an agent directly determine how much risk it can introduce (Source: Canadian Centre for Cyber Security). Analysts see the same pressure building. Gartner forecasts that by 2028, 25% of enterprise security breaches will be linked to AI agent abuse (Source: Gartner, Top Strategic Technology Trends for 2027). That is a forecast and not a count, but it shows where attention is heading.

The good news is that this is a familiar problem. The guidance says agent security belongs inside the security practices you already have, such as identity and access management and continuous monitoring, and does not call for an entirely new discipline (Source: Canadian Centre for Cyber Security). A job description is the same idea you already use for people, applied to software.

What goes in an agent's job description

Four things belong on it. Each one comes straight from the risks the agencies describe.

  1. A defined role. One agent has one job, and you can say what it is in a sentence. The guidance gives a clear picture of a role that is too wide: a calendar bot that can see all meeting data when it only needs the meetings of the person asking.
  2. Only the access that role needs. The guidance calls strict adherence to the principle of least privilege critical. Its second example of too much access is an email assistant that can write to any inbox, when it should reach one.
  3. Its own login. Agents sign in to other systems with secret keys or tokens, and criminals can steal them when the keys never change, are shared across several agents or are poorly protected. Give each agent its own identity and keep the keys managed.
  4. A manager who checks in. Review what the agent can reach on a schedule, and keep watching what it does. The guidance warns against checking permissions once at setup, because a stale approval can be used later to do things nobody intended.

(Source for all four: Canadian Centre for Cyber Security, Careful adoption of agentic AI)

What happens when an agent holds the master key

The agencies include a scenario that shows how this plays out. It is an example they wrote, not a reported incident. A company sets up an agent to handle purchase approvals and vendor emails. To reduce friction, it gives the agent broad access to its financial systems and contracts, and checks that access only on the day it goes live. Later, an attacker compromises a low-risk tool that the agent uses. The attacker inherits the agent's wide access and can change contracts and approve payments. Because the actions run under the trusted agent's identity, the logs look normal and the problem takes longer to spot (Source: Canadian Centre for Cyber Security).

Notice what failed in that story. The AI did not turn against the company. The access was wider than the job needed, and nobody went back to review it. A job description would have caught both.

Where to start

Start small. The agencies recommend beginning with low-risk, non-sensitive tasks and building up from there (Source: CISA, joint guide to secure adoption of agentic AI). Pick one agent, write its job description and give it the narrowest access that lets it do the work. Then put a date on the calendar to review that access.

If you would like a second set of eyes on what your agents can reach today, bring the question to a Qi Proof session. It is a free virtual session of 15, 30 or 60 minutes, built around what you want to see, with no contract or obligation.

Unlock a Qi Proof Session to run an Advisory Services consultation now: https://www.quickintel.com/contact/

QuickIntel Recent Blogs

Want to continue the conversation?
Book a consultation today.

Unlock a Qi Proof Session