How to run an AI agent incident drill for the IT team

IT administrators should drill what only they can contain: a faulty connection feeding agents the wrong data. Use a scenario in which a document store connection starts returning files from another client's folder. Over about two hours, the IT team practises pausing affected agents to draft-only, isolating the connection, escalating, recording each step and supporting the reporting decision.

What is IT's part in an agent incident?

Business teams pause agents and decide on customers. IT owns the plumbing: the connections that bring sources to agents, the identities and access rights agents use, the schedulers that run them, and the logs that sit alongside the record. When a connection fails or is misconfigured, every agent that reads from it is affected at once, across teams.

The IT operations manager owns the drill; the security lead often facilitates. IT administrators, a security analyst, the data protection officer and two business owners whose agents read from the connection take part. For the general format, see what an incident drill is.

The scenario

Monday, 09:00. A permissions change over the weekend gave the service account behind the document store connection access to a parent folder rather than one client's folder. Since Saturday, an onboarding agent and a claims agent, both reading from that connection, have occasionally drawn on documents belonging to other clients. The claims team spots a stranger's medical report attached to a draft.

The drill step by step

1. Person (facilitator): reads the scenario and starts the clock.
2. Person (claims team lead): pauses the claims agent to draft-only and calls the IT service desk. The test for IT: does the desk route it as a security incident, not a ticket?
3. Person (IT administrator): identifies every agent using the connection and asks each business owner to pause theirs to draft-only. Pausing is a business decision through the business's own limits; IT supports it.
4. Person (IT administrator): isolates the connection by reverting the permission change and, if needed, disabling the connection until it is checked. Agents that lose the source will hold or refuse actions that depend on it.
5. Person (security analyst): checks for compromise. Was the change a mistake or an intrusion? Review access logs alongside the record.
6. Person (IT administrator with business owners): lists affected drafts and actions from the record, which shows the sources each verdict used.
7. Person (DPO): assesses the breach and decides on notification under the data protection rules that apply.
8. Person (IT operations manager): decides on regulatory reporting of the ICT incident with compliance, where rules such as DORA in the EU financial sector or NIS2 for essential and important entities require it.
9. Person (IT administrator): records each step with time; the facilitator runs the debrief.

What does this drill teach IT specifically?

First, one connection can affect many teams. IT must know which agents read from which connections. If that map does not exist, the drill will show it.

Second, IT does not pause agents on the business's behalf by changing their bounds. Business owners tighten their own limits; IT changes connections and access. Keeping the two separate keeps the record clear about who decided what.

Third, the service account behind a connection needs the same discipline as any privileged identity. Least privilege, change approval and monitoring for scope changes are all IT controls, and the drill shows why.

Which controls does the drill test, and what evidence do they leave?

Routing. Evidence: the service desk ticket classified as a security incident, with time.

Pause by business owners. Evidence: limit changes in the record, each with a business owner's name.

Isolation. Evidence: the permission revert, with change reference and time.

Compromise check. Evidence: the security analyst's findings.

Reporting. Evidence: written decisions by the DPO and IT manager.

This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.

What should the IT operations manager measure?

• Time from the business alert to the connection being isolated.
• Time to identify every agent using the connection.
• Whether the security incident route was used from the start.
• Whether the permission change that caused it went through change approval.

Where do IT drills go wrong?

Ticket queue routing. If the service desk logs it as a routine request, hours are lost.

IT pauses agents itself. It feels faster, but it blurs accountability. Business owners should pause; IT should isolate.

No connection map. Without one, IT cannot answer the first question: which agents are affected?

IT supports every other team's drill; compare the operations team drill for capacity and the credit team drill for a data connection failure seen from the business side. Build your scenario with the AI agent incident drill script template, and see how to train a team to handle an AI agent incident and training teams for AI agents.

Where the Academy fits

Pryme Intelligence Academy's Foundations track gives IT staff the vocabulary of depth, verdicts and the record in about four hours. Administrators who also design agents and their connections may take the Builder track, and those leading regulated multi-agent deployments the Architect track, whose certificate is valid for 24 months.