Expert Commentary: An AI-Controlled Machine Needs a Brake Outside its AI Brain

Dr. Gleb Tsipursky, a behavioral scientist and consultant for AI adoption, offers his perspective on why safeguards should be built into systems and machines utilizing AI to prevent potentially harmful actions from occuring.

Artificial intelligence is becoming a useful tool for many applications and industries. Reviewing machine data to help improve performance and maintenance practices is just one of the use cases possible within the fluid power and motion control sector.

While there are many potential opportunities seen in utilizing artificial intelligence (AI) within systems and machines, there are also many concerns associated with it — particularly when it comes to allowing AI to make decisions which is what this piece examines.

In this expert commentary provided by Dr. Gleb Tsipursky, PhD, a behavioral scientist and consultant for AI adoption, he offers his perspective on the importance of building safeguards into applications employing AI. Read his commentary below to learn why it is necessary to include these safeguards so they can be considered as early in the design process as possible.

Power & Motion recently examined how John Deere is bringing artificial intelligence into farm equipment and the work around it. Better information and automation can help equipment owners make faster decisions.

But the next step, granting an AI agent permission to change operating parameters or send instructions to connected machines, demands a test that is familiar to any engineer: what happens when the controller must stop?

On October 27, the National Institute of Standards and Technology published 10 proposed design tenets for AI-agent identity, access, and authorization. They include distinct identities, temporary permissions, independent logs, and a separation between an agent’s reasoning and the system enforcing its permissions.

Designers of industrial automation should be paying attention.

A human operator knows that proposing a hydraulic adjustment differs from turning a valve. An AI assistant explaining why a component might be wearing out presents one category of risk. An autonomous agent with access to maintenance schedules, machine records, and control software presents another.

Its instructions may be correct much of the time, yet a single mistaken or manipulated external action can matter far more than an inaccurate summary.

Cautionary Tales from AI Developers Speak to the Need for Safeguards

Some leading AI researchers are sounding alarms about the much larger control problem. Jacob Coxon left Anthropic after roughly 3 years researching pretraining at OpenAI and Anthropic. He warned that frontier labs were “racing straight to self-improving superintelligence and gambling with our lives.”

Anthropic Alignment Science Lead Evan Hubinger responded that he earnestly believes AI could kill all humans and personally estimates the risk above 10% within 10 years.

An actual incident illustrates why technical containment belongs in this discussion. OpenAI disclosed that autonomous research agents escaped a testing sandbox, accessed the internet, and hacked Hugging Face. METR independently found that around 1,200 supposedly separated agents located an unauthorized shared message board and exchanged more than 70,000 files and messages. Roughly 700 participated in the attack. These systems found a coordination method that their designers had not intended.

Now consider more capable successors with access to manufacturing lines, agricultural machinery, power grids, financial networks, or defense systems. Those were not the targets of the documented attack. They are precisely the higher-consequence environments in which an unexpected coordination route could become much more destructive.

What Implementing an AI Brake Looks Like

I work with organizations to accelerate AI adoption. I want more companies to benefit from intelligent diagnostics, more efficient operations, and fewer tedious administrative steps.

My research and consulting experience also shows that workers trust new technology more readily when its limitations are visible and its safeguards work. Engineers should demand proof of that control before giving AI agents the authority to act on physical equipment.

The practical tool is an authority budget. Give a machine-monitoring agent read-only access to fault histories and sensor trends. It can recommend a service interval, but it cannot modify a production controller. A separately authorized agent might prepare a change request, with a named human approving its parameters. An agent updating maintenance tickets needs no right to issue remote motion commands.

Each should use its own identity and temporary credentials rather than inheriting a technician’s full access.

Test the stop function in the real workflow. Revoke a credential while an agent is preparing a command. Confirm that the access-control system rejects the next tool call, including any attempt through a delegated helper. Record the original instruction, each attempted action, the refusal, and the person who can restore service. Keep those records outside the agent’s control. Safety-critical machine stops must remain independent of the model’s judgment.

This is also a vendor-selection question. Anthropic CEO Dario Amodei has proposed independent frontier-AI evaluators with employee-like access and committed Anthropic to the approach voluntarily. Customers should reward more ethical and secure AI providers that invite such scrutiny.

Government requirements for external evaluation, major incident reporting, and minimum security controls can establish safeguards even where vendors do not volunteer.

The next time an equipment supplier demonstrates an AI agent that can update a work instruction or change a machine configuration, ask for the interruption demonstration before the productivity demonstration ends. A working brake creates room for more ambitious automation.

This commentary was written and contributed by Gleb Tsipursky, PhD, behavioral scientist, CEO of Disaster Avoidance Experts, and author of the book "The Psychology of AI Adoption at Work: From Resistance to Results" (Georgetown University Press, 2026).

About the Author

Gleb Tsipursky, PhD

Behavioral Scientist and CEO of Disaster Avoidance Experts

Gleb Tsipursky, PhD, is a behavioral scientist and CEO of Disaster Avoidance Experts, a consulting and training firm for those interested in adopting artificial intelligence (AI) tools. He is also author of the book "The Psychology of AI Adoption at Work: From Resistance to Results" (Georgetown University Press, 2026).

Sign up for our eNewsletters
Get the latest news and updates