The Loudest AI Warnings Are Coming From Inside the Industry
Joaquin "Jack" Fagundo, writing from Parkland, FL, on five AI warnings from this fall that business leaders and families shouldn't tune out.
I'VE SPENT 20+ years keeping other people's systems running, and over that time I've learned to listen for a particular kind of warning. Not the one shouted by critics, which is easy to wave off. The one that turns up quietly in a document a company would rather not have had to write.
This fall, those were everywhere. An AI lab put safety risk language into a filing meant for investors. The FBI started counting AI scams as their own category. Power prices began telling a story of their own. I made a short video on the five developments that stood out to me. This is the longer version, with what I'd actually do about each one.
When the builder writes the risk factor
On November 2, Anthropic, the company behind the Claude models, filed an updated document for its planned stock market listing. In it, the company said it has already used AI to help create its newer models, and that it may soon reach its own threshold for what it calls "dramatically accelerated progress attributable to AI." Researchers have a name for this, recursive self improvement: AI speeding up the work of building the next AI.
Anthropic says it hasn't reached that point, and it has pledged to keep humans in control of frontier development. It also wrote that using AI to build more capable AI "could also pose a considerable safety risk." An earlier version of the filing, first reported by Reuters in September, warned that advanced AI could pose "catastrophic or existential risks to humanity."
Companies don't put sentences like that in front of Wall Street casually. When a business with every reason to sell optimism writes a risk factor like this one, I read it twice.
Most companies will never build a frontier model. But nearly every company is now deploying AI that writes code, changes configurations and makes decisions inside real systems, and the lesson travels well. If an AI tool can touch production, a person should approve what it does, someone should be able to see exactly what it changed, and someone should be able to switch it off.
If an AI tool can touch production, a person should approve what it does.
Worms, and the devices nobody owns
In October, the same company published research on AI worms. Picture a small AI model that doesn't just assist an attacker but spreads on its own, moving from one poorly secured device to the next. Researchers ran a model across 15 off brand Android TV boxes and tested whether it could find a device, take control, set itself up and keep going. Their conclusion was that small models may be months away from copying themselves this way.
Some perspective matters here. The tests used deliberately weak, emulated devices, and nothing has spread in the wild. Still, I've never walked a corporate campus or a data center that didn't have a long tail of devices nobody really owned: cameras, badge readers, the smart TV in the conference room that facilities installed and IT never heard about. That's exactly where a risk like this would live.
The fix isn't glamorous. Know every device on your network, including the ones facilities bought. Put smart gadgets on their own network, at the office and at home. Retire anything that no longer gets firmware updates. And skip the no name streaming boxes and unofficial app stores, which is advice that belongs at the kitchen table as much as in the server room.
The scam now has a dollar figure
For the first time, the FBI's annual Internet Crime Report broke out AI as its own category. Americans filed 22,364 complaints involving AI in 2025, with about $893 million in reported losses. The schemes themselves are old: business email compromise, romance scams, fake job offers, bogus investment pitches. What's new is the cloned voice on the phone, the convincing video, and the polish.
What works against this is process, not more technology. In a business, nobody changes payment details without a callback to a number already on file, and wire transfers need two approvers. Finance teams should learn to treat urgency itself as the red flag. At home, agree on a family code word, and hang up on anyone who pressures you for money, however familiar they sound.
The bill nobody voted on
On PJM, the grid that serves about 67 million people across 13 states, wholesale electricity prices in early 2026 averaged about $137 per megawatt hour, roughly 76% higher than a year earlier. The grid's independent market monitor named data center load growth as the primary reason for tight capacity. In a Gallup poll, 71% of Americans said they oppose a data center being built in their area.
I've argued in my piece on AI data center spending that budgets in this industry now start with megawatts instead of square footage. This is the other side of that coin. Data centers are going up faster than power plants and transmission lines, and households end up sharing the cost. If you run a business, expect your energy costs to rise, and get power delivery dates in writing from your colocation and cloud partners before you commit to an AI program. If you own a home, your utility's rate cases are worth following now.
The money for the job
In April, AI was the leading reason companies gave for planned job cuts in the United States for the second month in a row, according to Challenger, Gray & Christmas. Through April, AI was cited for 49,135 planned cuts this year. Andy Challenger put it bluntly: whether or not people are being replaced by AI, the money for those roles is.
I've been through enough transformation programs to have a view on this. The companies that handle it well put part of the savings back into retraining the people closest to the work. Those people know where the processes break, and they are usually the ones who end up making the AI useful.
Where that leaves us
None of these warnings calls for panic, and none of them needs a breakthrough to address. They call for discipline: a human in the loop on systems that matter, an honest inventory of what's connected, a callback before money moves, power plans in writing, and a retraining budget funded from the savings. That's the same discipline infrastructure work has always asked for. The stakes are just higher now.
Joaquin "Jack" Fagundo writes about data centers, cloud and the business of AI from Parkland, FL. He has spent 20+ years leading enterprise infrastructure and advises companies through JFagundo Consulting Services. Read more about Jack or find him on LinkedIn.
Sources
