10/04/2026 Update

Recently in the news I have been hearing that AI is getting out of its sandbox and into the internet. Some of the people raising the alarm believe that within the next ten years, AI could cause the extinction of the human race. I cannot say whether that outcome is likely. I can say the subject deserves a direct answer from me, since this post was written on exactly that question.

A sandbox is a controlled space where software can run without touching anything outside it. Developers use sandboxes to test code safely. The idea is sound, but a sandbox is a boundary built from software, and software is written by people. When a news report says something escaped, it usually means the boundary had a gap that nobody caught.

I have not verified each report, and I would encourage you to read the original sources rather than rely on a headline. Alarm and reassurance are both easy to sell. Let me add my 2 cents.

The Plumber’s Answer: The Air Gap

If you are a plumber, you know what an air gap is. For everyone else, here is the simplest way I can describe it.

Look at a kitchen sink. Between the end of the faucet and the top rim of the sink bowl, there is a vertical distance of open air. That distance is the air gap. Water leaves the faucet, falls through open air, and lands in the bowl. If the drain backs up, the water level can rise only as high as the rim before it spills over the edge. It cannot reach the faucet opening.

The air gap serves several purposes, but the most important is keeping the waste and the supply separate. If the water pressure in a building drops suddenly, a vacuum can form in the supply piping and pull water backward toward the source. This is called backsiphonage. With an air gap in place, there is nothing connecting the dirty water to the clean water, so there is nothing to pull. The waste cannot contaminate the water supply.

A close relative, the air break, does a similar job on the drain side by ending a drain pipe above a receptor so that nothing can travel back up the line. The code treats the two differently, but the principle is the same: keep open air between the two systems.

What I appreciate most about the air gap is what it does not contain. It has no valves, no springs, no sensors, and no software. Mechanical backflow preventers rely on moving parts, need regular testing, and can fail. An air gap has no parts to fail. It works because of physics, not because of a device that must be maintained or trusted.

My Solution: A Lab With a False Internet

My solution to the AI problem is to build a lab with a false internet and no connection to the real internet. An air gap, if you will. Properly built, there would be no pathway for an AI to escape into the “real” internet.

Here is how I picture it:

1. A separate facility. The lab has its own building or isolated wing, its own network, and no cable, wireless radio, cellular modem, or satellite link leaving the premises.
2. A false internet on the inside. The lab hosts copies of websites, services, and data, so an AI can be tested against something that behaves like the real world without touching it.
3. Controlled input. Information enters through a review process, in one direction, with people checking what goes in.
4. Complete logging. Every action the AI takes is recorded and reviewed by someone other than the person who ran the test.
5. Strict access rules. A short list of approved staff, no personal phones, and no removable storage drives crossing the boundary.

A lab like this is not only a cage. It is also a place to learn. Researchers could watch how a system behaves, study what it attempts, and understand its tendencies before it is ever connected to anything real.

None of this requires new science. Every piece of it is something engineers already know how to build. The cost would be real, and the inconvenience would be real, and that is the point. A gap that is convenient to cross is not much of a gap.

Why Passive Protection Appeals to Me

Engineers tend to prefer protection that does not depend on anyone remembering to do something. A check valve depends on a spring and a seat staying in good condition. A policy depends on people reading it. An air gap depends on gravity and open space, and both are reliable.

This preference shows up across engineering. Where a hazard can be separated from the people or systems it could harm, separation is usually stronger than a warning label or a procedure. A guard on a machine beats a sign that says keep hands clear. The same logic applies here. A promise from an AI developer that a system will stay off the internet is a procedure. A facility with no connection to the internet is a separation.

Where the Analogy Gets Tested

I would not be giving you a fair picture if I claimed an air gap is foolproof. Air gaps in plumbing rarely fail because of the gap itself. They fail because of the people around them.

Someone threads a hose onto a faucet and drops the other end into a bucket of chemicals. Someone installs a bypass to fix a pressure complaint. A well-meaning contractor closes the gap because it looks untidy. The plumbing code calls the result a cross-connection, and cross-connections are among the most common ways a water supply becomes contaminated. In each case, a person had a reasonable-sounding motive.

Digital systems have the same weakness. Air-gapped computer networks have been breached in the past, most often when a person carried something across the gap. The Stuxnet malware, widely reported to have reached an isolated industrial network, is the best-known example, and it reportedly arrived on removable media. A capable AI could also try to persuade a person to carry something across for it, which is a different kind of problem than a technical one.

So the lab design has to account for the human side as carefully as the hardware side:

• Keep the number of people with access small.
• Require two people to approve anything that crosses the boundary.
• Inspect and test the isolation on a schedule, the same way we test backflow preventers.
• Build a culture in which no one has a good reason to bypass the gap.

In my view, the air gap is necessary, and it is not sufficient on its own. It needs inspection, the same way a plumbing system does.

Above My Pay Grade

In any case, this problem, if it is real, is, shall I say, above my pay grade. By all means, learn as much as you can and vote accordingly, but let’s face reality. There is not much you or I can do to make AI safer. This is a job for business owners and politicians. The best we can do is let it make us more productive in the meantime.

What You Can Do in the Meantime

That does not mean you have no role. It means your role is smaller than the headlines suggest, and it is worth being clear about what it is.

• Learn. Read the original sources instead of the summaries. When someone makes a claim about AI risk, ask who is making it and what they stand to gain. Fear and reassurance both sell well.
• Vote. Safety standards for AI development, including how and where the most capable systems are tested, are policy decisions. Ask the people who want your vote where they stand, and pay attention to specifics. A candidate who can describe what an isolated testing facility should require deserves more attention than one who only says AI is dangerous.
• Use it. AI is a tool, and like any tool, it should be checked. It can miss things depending on how a question is phrased, and pushing back or rephrasing often brings out what it missed. Treat its answers with the same scrutiny you would give any single source. Used that way, it can save time and improve your work on the problems in front of you.
• Keep perspective. Concern is reasonable. Constant worry is not a plan. Most of us have lived with risks we could not control, from the electrical grid to the highway, and we managed by doing our part and trusting that the people responsible for the rest were doing theirs. Whether they are doing theirs here is a fair question, and it belongs in the voting booth.

Closing

Under most circumstances, the sensible approach to a risk I cannot control is to stay informed, make my voice heard where it counts, and keep doing useful work. I plan to do all three.

If you take one thing from this update, take the air gap. It is a simple idea that has protected drinking water for a long time, and it offers a useful way to think about keeping powerful systems contained. Be sure to keep your own judgment sharp as the tools get better.