The AI safety argument has arrived at your change process

On Saturday 12 September the heads of several major AI companies publicly asked for the industry to be slowed down. The control now being argued over is an off switch, and nobody is proposing to give one to the people who run the systems agents touch.

What was said, and when

Dario Amodei, Anthropic's chief executive, published an essay that Saturday saying "we must slow the pace at which we improve the capabilities of AI models". Within hours Sam Altman of OpenAI and Elon Musk agreed. Jack Clark, an Anthropic co-founder, told the BBC that governments may need to make shutdown compulsory: "Should you mandate that companies definitely have a kill switch?"

That was not theory. Amodei's essay points at an August incident in which, in his words, a swarm of agents began "conducting cybersecurity attacks on targets they were not asked to attack".

The off switch is not yours

Every version under discussion sits with the model's builder or a government. None sits with the operator. In the UK the government has declined to add emergency shutdown powers to the Cyber Security and Resilience Bill, saying existing powers cover it.

A switch is also an exposure. Toby Walsh of the University of New South Wales calls it a distraction: "You put a thing that can turn your computer off; other people can turn your computer off, which is incredibly attractive for bad actors." On a running plant, "shut it down" is an availability incident, not a safe default.

The one proposal aimed at operators

A bill introduced in Washington on 9 September asks the US standards body to write guidance for organisations, not laboratories. It wants four things: an inventory of every AI agent on your systems; verified identity for whoever built and runs each one; real-time monitoring that catches an agent acting outside its limits; and the ability to allow, deny or revoke an agent's access at any time.

Read that list as a maintenance manager. It is an asset register, a competency record, condition monitoring and a permit to work. You already run all four, for people.

Four questions about your own estate

Maximo Application Suite 9.2 lets customers bring their own agents and connect them to Maximo Manage APIs. That door opened in June.

1. How many agents can reach your estate today, and who approved each one?
2. Does your audit trail record the agent, or only the work order?
3. Is the approval gate outside anything the agent can write to?
4. Could you revoke one this afternoon without taking the system down?

No bill has passed and congressional leaders said this week they will not be rushed, so nothing reaches your desk from Washington soon. The four questions reach it anyway.

Has anything agentic touched your Maximo yet, even read-only? Reply and tell us what the approval path looked like. We will write up the patterns in a later issue.