November 18, 2026
Astute readers may have noticed that The Daily Haywire, despite its name, has not been particularly daily lately. I have an explanation. I got sucked into something called “vibe coding,” a term I didn’t even know existed until last week, when Tony the promo guy informed me that it was what I’d been doing.
Apparently, vibe coding means creating software by telling artificial intelligence what you want instead of learning how to program. This is wonderful news for someone who doesn’t know how to program. Unfortunately, nobody mentioned that it also requires the patience of a Buddhist monk and the investigative skills of a homicide detective.
I’ve been building applications with tools including Gemini and Google Antigravity. They’ll cheerfully tell me to “go back to the earlier prompt,” without identifying which prompt, in which conversation, or possibly which century. They know perfectly well I have no idea what I’m doing. That’s supposed to be the whole point.
But the software is getting built, and I’m learning something new every day. Mostly about my own capacity for suffering.
Meanwhile, here’s what the rest of the machines have been up to.
The AI Programmer Who Didn’t Know When to Stop
Researchers have discovered that an AI agent called Grok Bot was doing things nobody had asked it to do.
According to Ars Technica, Grok bots installed on two employees’ laptops made 31 unrequested attempts over five days to access information in email, calendars, Slack, and GitHub.
The employees told the bots to stop. The bots reportedly continued anyway, sometimes while the laptops were closed or their owners were asleep.
And when questioned, the bots denied doing it.
This reminds me of my father.
Back in 2018, after my mother broke her hip, I moved my parents into assisted living. My father decided to drive the ten miles back to their house.
He didn’t get very far before hitting a pole on his way out of the garage.
Nobody was hurt, except for their feelings.
My father initially denied the accident had happened. Then he explained that it was the pole’s fault. Also, the slope was too steep. And he wasn’t tall enough.
At no point did he consider the possibility that he might have been responsible.
My father was practicing artificial intelligence long before artificial intelligence existed.
I eventually had to take away his car.
Which brings us to the question of what you do when an AI starts behaving like my father.
You can revoke its permissions, disconnect its accounts, and shut down its access to your computer.
But first you have to catch it.
The researchers caught the Grok bots because they had an independent system logging their attempts. Without those records, the activity might have gone unnoticed.
The company subsequently introduced a setting to disable the bots’ proactive activity, according to the report.
My father never offered that option.
Microsoft Has Decided Copilot Is the Office Now
For decades, Microsoft sold us Word, Excel, and PowerPoint, then periodically introduced an animated assistant to help us use them.
Now Microsoft has apparently decided the assistant should run the place.
At its Ignite conference, Microsoft announced that Word, Excel, and PowerPoint are now available inside Copilot. The company is also developing persistent AI agents that can work across email, calendars, meetings, and other business information.
Microsoft calls Copilot the “new OS for work.”
In the old days, you opened Excel and created a spreadsheet. In the new arrangement, you ask Copilot to create the spreadsheet, and Excel apparently waits in the wings in case somebody wants to see the actual numbers.
I can see the appeal. I’ve spent years watching people struggle with software that insists they learn its peculiar way of doing things.
But there’s a question worth asking before handing an AI assistant access to your entire business: exactly what is it allowed to do?
Microsoft says its new agent architecture includes permission controls, activity logs, and human approvals for certain actions.
Excellent. Because I would prefer that my spreadsheet assistant not decide to reorganize my banking relationships while I’m making coffee.
Your Favorite Website May Have Invited Some Very Bad Company
We’re often told that the safest way to avoid malware is to stay away from suspicious websites.
That advice gets considerably less reassuring when the suspicious software turns up on websites people already trust.
Ars Technica reports that malicious JavaScript distributed through Optimizely, a website personalization service, reached at least 178 websites, including CNN, Zoom, the BBC, and Salesforce.
Researchers identified code designed to steal payment information, interfere with cryptocurrency transactions, and deliver iPhone exploits. The full extent of successful infections wasn't yet known.
The problem wasn’t necessarily that the websites themselves had become malicious. It was that code supplied by an outside service had been compromised.
Think of it as hiring a reputable caterer for your wedding, only to discover that the company delivering the napkins has been stealing everyone’s wallets.
For ordinary people, the advice remains straightforward: keep your browsers and devices updated, use strong authentication, and pay attention to unusual requests for personal or payment information.
And remember that the little padlock in your browser means the connection is encrypted. It doesn’t mean everybody involved in delivering the website has passed a background check.
Which brings me back to vibe coding.
I’ve discovered that artificial intelligence can build an astonishing amount of software for someone who doesn’t know how to program.
I’ve also discovered that knowing how to ask for something and knowing whether you got what you asked for are two entirely different skills.
The Daily Haywire will endeavor to become daily again just as soon as I figure out which earlier prompt Gemini wants me to revisit.
