How to Stop an AI Agent That Goes Wrong: Access Limits, Alarms and an Off Switch

Matthias Heim · 2026-11-16

Decide before an incident who may switch off which system, give agents the least access that works, and have them stopped automatically when they touch what they shouldn't. With a checklist for regulated companies.

Name one person who may stop an AI agent, or the system it works in, and agree the conditions for stopping it before the agent gets access to real systems. Then limit what each agent can reach, and have it stopped automatically when it touches something it has no reason to touch.

A simple test: who in your company may switch off the webshop at 2 a.m. without asking anyone? If nobody can answer, that decision is missing, and an agent will not make it for you.

This article covers both, with the safeguards we use for our own agents and a checklist you can copy.

In short

Decide in advance who may take which system offline, under which conditions, and who informs customers and the board.

Give agents the least access that works. In our setup, agents can draft emails but cannot send them.

Add alarms and an automatic stop: decoy passwords that no real task uses, and an off switch that agents can trigger but only a person can reset.

Why Watching Is Not Enough

An AI agent is software that works toward a goal on its own, using the access you give it: it reads email, updates the CRM, changes files. That is the point of it. The risk is that it acts with the same access when the input is wrong.

The wrong input is usually ordinary: an email with hidden instructions, a web page, a software package that was tampered with before it reached you. The agent does not need bad intentions to cause damage. Access and the wrong input are enough.

Many early setups rely on someone reading logs or reacting to an alert. That works with two agents. It does not work when agents run day and night and the person on watch has other work. At Lailix, 64 scheduled automations were running on agents at the end of October 2026. Nobody watches each step, so our protection has to work without anyone watching.

Decide Before the Incident

Most of the delay in an incident happens before anyone touches a system. Three questions decide how fast you can act:

Where do we run this? Including connections such as an order portal that pulls stock levels from the ERP.

Who may take a system offline, under which conditions, without calling a meeting?

Who tells customers, and who tells the board?

Your security provider can tell you that something is wrong. It usually cannot decide to switch off your shop. If you answer these questions during an incident, you lose the first day.

Keeping the answer to the first question current is a good job for an agent. It can maintain the list of systems and connections and, when a new security warning appears, answer "are we affected, and where?" in minutes. The decision to switch something off stays with a person.

Five Safeguards We Use for Our Own Agents

For the agents themselves, we built safeguards that do not depend on anyone noticing a problem:

Least access. Each agent gets only the access its task needs. Our agents can draft any email. None of them can send one.

Decoys. Fake passwords and files that no legitimate task ever uses. If an agent touches one, that is the alarm.

Tamper alarms. An attempt to change the safeguards themselves, or to switch one off, also raises the alarm.

Automatic stop. It is built so that when the alarm goes, every agent stops within seconds and loses its access, without waiting for someone to read an alert.

Only people restart. Agents can trigger the stop but cannot undo it. Restarting takes a person at the machine, confirmed with a fingerprint.

The off switch works in one direction: agents can trigger it, only a person can reset it.

Before installing it, we had other AI models attack the setup. One of them found a way an agent could have turned the off switch against us. We closed that gap.

We also switch it on in stages. For the first weeks it only records what it would have stopped. It stops agents for real only after a clean trial period and a drill, because an off switch that fires by mistake also costs trust.

The Same Rule in Client Projects

In client projects we apply the same rule: the agent proposes, a person decides. An agent may read tickets, documents or records and draft a reply, a change or a correction. A named person reviews the result before anything reaches a customer or a live system.

Checklist for Regulated Companies

If you work under Swiss data protection law or industry supervision, add these points to your setup:

A named owner for every agent, plus a deputy, so the knowledge does not leave with one person.

A record of what each agent may access, kept together with your record of processing activities under the Swiss Federal Act on Data Protection, FADP (nDSG).

An audit trail: which agent did what, with which input, and who approved it.

A tested stop: who can stop which agent, how, and how you restart it. Run a drill before you rely on it.

An incident plan that names who informs customers, the board and, if a data breach is likely to put the people affected at high risk, the Federal Data Protection and Information Commissioner (FDPIC).

We set these up with teams as part of how we work , starting with the first workflow.

Frequently Asked Questions

What are AI agent guardrails?
Technical limits on what an agent can do, plus automatic responses when it does something unexpected: restricted access, decoys and tamper alarms, and an off switch that stops agents without waiting for a person to react.
What is a kill switch for AI agents?
A mechanism that stops running agents and cuts their access. Ours is built to trigger automatically when an agent touches a decoy, and only a person at the machine can turn the agents back on.
Who should decide to take a system offline during a security incident?
A named person, with conditions agreed in advance, so the decision does not wait for a meeting. Your security provider can raise the alarm, but the business decision belongs inside the company.

How we work · Work · Insights