When AI can change factory settings or influence other safety-critical operations, ordinary cybersecurity controls are not enough, argues Modcon Systems chief executive Gregory Shahnovsky. His proposal is to put independent safeguards between an AI system and the equipment it could affect, rather than asking the AI to judge whether its own actions are safe.
Watch Desk analysis
What happened
In a Forbes Technology Council article published on 13 October, Shahnovsky argues that AI systems should be assessed according to the possible consequences of failure. A tool that summarises documents is not the same proposition as one that can change industrial operating conditions, he says.
For high-consequence uses, his recommended measures include defined operating limits, restricted permissions, required human approval and comprehensive logging. He also argues for a separate protection layer able to reject an out-of-bounds command, remove write access, isolate the AI or return control to an operator. The safeguards, he says, should still work if the AI is compromised.
Why it matters
The practical distinction is between an AI system that makes a suggestion and one with authority to affect equipment, people or infrastructure. If a cyberattack or faulty decision can reach the physical process, cybersecurity failure can become a safety failure too.
Shahnovsky points to industrial safety engineering, including hazard and operability reviews, as a model for asking what could go wrong, what the consequences would be and which protections would stop the harm. He cautions against copying industrial safety labels wholesale into AI governance; the useful principle is matching the strength and independence of safeguards to the stakes.
Our read
This is a useful argument from an executive whose company works in industrial analytics and AI optimisation, not evidence that a particular AI deployment has failed. Its strongest point is architectural: do not make the system being controlled the sole judge of whether its own actions are acceptable. For AI with real-world authority, a separate brake is not bureaucratic fuss. It is the part that matters when everything else goes wrong.
What to watch
- Whether organisations assess AI systems by what they can control, not only by what model they use.
- Whether high-consequence deployments include safeguards independent of the AI itself.
- Whether regulators or standards bodies turn these principles into concrete requirements.
Discussion spark: For AI that can affect physical equipment, should independent safety controls be mandatory, or should that depend on each organisation’s assessment of the risk?
Sources and evidence
- AI Security Should Learn From Industrial Functional Safety – Forbes (13 October 2026, 13:30 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.