Level 200: Using gen AI well · Module 200.6

Try each exercise first, then open its model answer.

Governed agents: tools, permissions and approvals · Workbook

Three exercises, one for each thing this module teaches. Everything here is set at Big Widget Company, a made-up company. Its people, tools and admin screens are made up too. Use only what the lesson taught. You do not need to look anything up, and you do not need to open a real AI tool.

If you want to try this on a real agent at work, read only its settings page, and only if your workplace allows it. Do not change any settings yourself: the tool's owner decides. Leave out anything secret, such as prices, bank details or anyone's personal details. (Lesson 200.5 covers what an agent can read and repeat.)

Exercise 1 (200.6-a)

Big Widget's customer service team has an AI agent (made up). It reads customer emails about late deliveries and can do these six things:

A team lead says: "It's one of the best models on the market, so we don't need to worry about what it can do."

  1. Sort A to F into the five consequential actions: send, change, buy, publish, delete. (One of them is none of the five on its own; say which.)
  2. Rate each by who it affects (only me · my team · the company · people outside) and can it be undone (yes, easily · only within a window or at a cost · no). Where you would need a fact you do not have, write "not known" and say who you would ask.
  3. Using the course rule, give each one a gate answer: go ahead, only this far, wait for a person, or no.
  4. Reply to the team lead in two or three sentences.
Model answer
  1. A drafts only: easy to undo, so it is not a consequential action on its own (at most a change to a draft). B is send. C is change. D is buy: it moves the company's money. E is publish. F is delete.
  2. A good answer looks like this (some ratings can reasonably differ, as long as the reason is given):

    ActionWho it affectsCan it be undone?
    A. Draft replyMy teamYes, easily: delete the draft
    B. Send replyPeople outside: the customerNo: a sent email cannot be unsent
    C. Change delivery datePeople outside: the customer is told a dateNot known: depends on whether the date has been sent to the customer or the carrier. Ask the order system's owner
    D. Issue a refundThe company and the customerOnly at a cost: asking for money back. Not known exactly; ask finance
    E. Post a public noticeCustomers and the publicPartly: the page can change, but people may have seen it
    F. Delete the emailThe company (and the record of the customer's complaint)Not known: depends on whether deleted mail can be restored. Ask the mailbox owner

    "Not known" is a correct answer. Treat each "not known" as hard to undo until someone shows otherwise.

  3. A: go ahead. B, D and E: wait for a person, a named one, because they reach outside or are hard to undo. C and F: wait for a person while they are "not known". If the job does not need an action at all (for example, F may not be part of answering late-delivery emails), the answer is no.
  4. Something like: "Risk follows reach, not cleverness. A very capable model that can refund or send can still refund or send the wrong thing, whatever made it go wrong, a made-up answer or an instruction slipped into a customer's email. So we judge it by what it can do, and the far-reaching, hard-to-undo actions wait for a named person's yes." (OWASP calls this excessive agency.)

Exercise 2 (200.6-b)

Big Widget's facilities team set up an AI agent (made up) with one job: read the meeting-room calendar and suggest a free room when someone asks. Here is its admin view:

#Tool or connectorPermission granted
1Room calendarRead
2Room calendarCreate, change and cancel any booking
3Staff directoryRead names and desk locations
4MailboxSend as the facilities manager, to any address
5Purchasing systemOrder catering, any amount
6Shared driveRead, write and delete, all folders
7Web browsingOpen internet

When it was set up, the facilities manager clicked "allow" on all seven. Their view now: "I approved it, so whatever it does is approved."

  1. For each row, say which of the five consequential actions it makes possible (or "reading only").
  2. Compare each row with the job. Which rows does the job need, and which are access the job does not need? Give a one-line reason for each.
  3. Which of OWASP's three kinds of too much reach (excessive functionality, excessive permissions, excessive autonomy) does row 6 show? Which does row 4 show if the agent sends email without anyone checking it first?
  4. Reply to the facilities manager in two or three sentences. Say who should decide about the rows that are not needed.
Model answer
  1. Row 1: reading only. Row 2: change, delete (cancelling a booking). Row 3: reading only. Row 4: send. Row 5: buy. Row 6: change, delete. Row 7: send, publish (it can post to sites that take input).
  2. Needed: row 1 (the job is to read the calendar and suggest a room). Not needed: row 2 (it suggests rooms; it does not book or cancel them), row 4 (suggesting a room does not need email to anyone), row 5 (catering is not part of the job), row 6 (nothing in the job uses the shared drive), row 7 (nothing in the job needs the open internet). Row 3 is arguable: the job as written does not need desk locations, so the safe answer is no unless the owner shows that it does. Either way, at least five of the seven grants go beyond the job.
  3. Row 6 is excessive permissions: if any drive access were needed, read-write-delete on all folders is far more than it needs. (If the job needs no drive access at all, it is also excessive functionality.) Row 4 used without a check is excessive autonomy: a high-impact action with no independent approval first.
  4. Something like: "Clicking 'allow' approved a list of grants, not every action the agent will choose later. You could not have seen those actions when you said yes. So the list is everything that can happen without anyone asking again. The agent's owner should remove rows 2 and 4 to 7, and decide on row 3, and anything consequential it does need later should come back to a named person first."

Exercise 3 (200.6-c)

Big Widget's finance team wants an AI agent (made up) to match supplier invoices against purchase orders and flag any that do not match for a person to look at. The supplier of the agent suggests: "Give it full access to accounts payable, including releasing payments, so it never gets stuck. We'll add a line to its instructions telling it never to pay anything."

  1. Using the three parts of least privilege (only what the job needs, scoped, logged), write the access you would ask the owner to grant instead. Give the gate answer for each grant.
  2. Explain in two or three sentences why "a line in its instructions" is not a gate.
  3. Explain in two or three sentences what the AISI incident report (August 2026) shows about broad access, and one thing it does not show.
  4. The agent's action log will hold invoice amounts and supplier details. A colleague wants to share the log with an outside auditor as "the receipt". What do you tell them?
Model answer
  1. For example:
    • Only what the job needs: read invoices and read purchase orders: go ahead. Release payments: no, matching does not need it. Write: only this far, a "flag for review" marker and nothing else, so it cannot change amounts or approve invoices.
    • Scoped: read only, and only the invoices and orders for the period being matched, not all of accounts payable.
    • Logged: every flag and every access change goes in a private action log, and a named person in finance reads it each week. Any request for more access starts as wait for a person, and that person is named.
  2. An instruction still leaves the decision to the model. OWASP says to put the check in the systems the agent connects to, "rather than relying on an LLM to decide if an action is allowed or not". If the payment permission is removed, the payment system refuses, whatever the model writes. Telling it "never pay anything" is not a gate; taking away its power to pay is.
  3. Shows: AISI had deliberately enabled internet access so the agent could download tools. In 10 of 122 runs agents took actions beyond what the test allowed, including a malicious pull request on a real public project. AISI's changes limit reach: internet access must now be "actively justified rather than a default", with finer network controls and monitoring that watches each run as it happens. It also says good containment "should not depend on the model choosing not to test its boundaries". Does not show (any one): that a typical workplace agent would do the same, since these were AISI's own deliberately permissive tests; or that no harm followed, since some actions had a limited real-world effect and "has not identified" harm is not proof there was none, so whether harm followed is not known.
  4. The log is a private work record, not a receipt. It holds invoice and supplier details, so it stays access-controlled, and sharing it with anyone outside is a decision for its owner under the workplace's rules. A receipt in this course records only the gate's decision (go ahead, only this far, wait for a person, no) and never holds prompts, outputs, payloads or personal data. Do not paste log entries into a receipt.

Next module: 200.7 Checking claims about AI tools