Least privilege for AI agents: write the job description first
· By Bee Flow · 6 min read
Six national cyber security agencies have published joint advice for AI agents that use tools: only the access the task needs, a person before anything with impact, and a log you can read. What that looks like in a small organisation, with a made-up supplier-mail agent.

On 1 June the OWASP GenAI Security Project published version 2.01 of its State of Agentic AI Security and Governance report, its guide to securing and governing AI systems that act on their own. It comes a month after six national cyber security agencies, among them CISA, the NSA and the UK's NCSC, published joint guidance called Careful adoption of agentic AI services on 1 May. On 15 May the NCSC summarised that guidance in a blog with the subtitle "make sure you can walk before you run".
The joint guidance was written for government and critical infrastructure, but its core advice fits a company of twenty people just as well. This post turns the part about privileges into something an IT lead can set up this month.
What the agencies ask
The guidance is plain about where the risk sits: "Privilege risks are a key concern for agentic AI and strict adherence to the principle of least privilege is critical." The authors recommend never granting an agent "broad or unrestricted access, especially to sensitive data or critical systems", and using agentic AI only for low-risk and non-sensitive tasks.
Their own example is close to home for many offices. An organisation lets an agent handle procurement approvals and supplier communication, and gives it broad access to financial systems, email and contract repositories, "evaluating permissions only at initial deployment". An attacker compromises one low-risk tool in the agent's workflow, inherits those privileges, modifies contracts and approves payments. Because it all happens under the agent's identity, the audit logs look legitimate.
Four practices from the guidance carry most of the weight for a small organisation:
- limit entitlements "to the exact resources, operations and timeframes needed", and replace long-lived secrets with credentials that expire when the job is complete
- decide in advance when a person must approve, and leave that decision to the system designers or operators, "not delegated to the agentic AI system"
- log the tools an agent uses, with the results captured "in a human-readable format"
- do not let agents modify their own privileges
The NCSC blog adds the human side. It asks you to give agents "only the minimum access they need, for the shortest time required", and to settle "who owns an agentic system, who approves its access, who monitors its behaviour, who reviews incidents, and ultimately who can stop it" before the agent is connected to real systems. Its test is short: "If you cannot understand, monitor or contain an agent's actions, it is not ready for deployment."
An agent is a colleague with a job description
That list reads like something you already do for people. A new colleague in purchasing gets access to the purchasing mail, not to the payroll folder. On their first day they may draft a reply to a supplier, but a change to payment terms goes past their manager. And when something goes wrong, the mailbox shows what was sent.
Write the same four lines for an agent before you build it. Which systems it may read. Which it may change, and how. When a person must approve. How you see afterwards what it did. Then add a fifth line with the owner's name: the person who can switch it off.
A made-up example: the supplier-mail agent
Groothandel Voorbeeld is a made-up wholesaler with 25 people. Two buyers handle about forty supplier mails a day: changed delivery dates, price updates, order confirmations, the occasional complaint. The IT lead wants an agent that reads those mails and prepares replies.
On day one the job description is short.
| Day one | Later, once it has earned it | |
|---|---|---|
| Reads | the buyers' supplier mail, a knowledge base with the purchase conditions | the list of open orders, read-only |
| Changes | creates a draft reply, nothing else | still drafts only, now also from a scheduled run every morning |
| Approval | the buyer reads every draft and sends it from their own mailbox | an approval step before anything leaves the building |
| Afterwards | every tool call visible in the conversation | a run history with each step's input and output |
| Never | send, delete, forward, payment details, supplier bank accounts | the same |
The agent gets three tools on day one: search mail, read a message, create a draft. It does not get the tool that sends. It is shared with the two buyers and nobody else, so it reaches only what their own accounts reach. A buyer opens the draft, fixes what needs fixing and presses send. That is the human approval, and it takes about ten seconds per mail.
The last row matters most. A supplier mail is text from outside your organisation, and text from outside can carry instructions meant for the agent rather than for you. With sending, forwarding and bank details out of the agent's reach, a mail like that can at worst produce an odd draft, and a buyer will notice it.

What you add later, and when
Add rights one at a time, and only when you can name the evidence. After a month of drafts that the buyers rarely rewrite, read access to the open orders lets the agent quote the right order number. That is still a read. Write down the date and the reason, as you would for a colleague's access request.
The next step takes the agent out of the chat window and into a scheduled job that drafts replies before the buyers come in. Nobody is watching it run, so the guidance's advice on approval points now applies in full. Put an approval step in front of anything that sends or writes, and test the run without side effects first.
Some rights never come. A change to a supplier's bank account, a payment, a contract amendment: these stay with people, however good the agent's record. Write that list down now, while it is still short.
Review the job description every quarter. Rights that were added for one project and not used since go back. That is what "for the shortest time required" looks like in a small office.
How Bee Flow handles this
In Bee Flow every agent has an explicit set of tools, picked one by one in the Agent designer, with a deny-list for tools it should never call, and you can share it with specific groups only. Integrations are switched on per organisation and per group, and an agent works through the connected account of the person using it, so it cannot reach more than they can. In the Gmail integration a draft and sending are separate tools, and Outlook also comes in a read-only variant, so the day-one column above is a matter of ticking boxes. Every tool call shows up in the conversation with its full request and response. Automations add approval steps that pause a run until a person decides, a dry run that does not invoke tools which send, create, update or delete, and a run history with each step's input and output. If you would rather have us build such an agent or automation with these limits, see automation as a service. More on Privacy & Security, or get in touch to try the supplier-mail example on your own mail.