The AI Gold Rush Is Breaking Everything—Including How We Work

---

The ex-Meta employee who downloaded 30,000 private Facebook photos didn’t need to hack anything. He just had access. That’s the detail that should keep you up at night—not because it’s shocking, but because it’s boring. It’s the kind of insider threat that happens when a company scales AI capabilities faster than it can manage the humans touching the systems.

Meanwhile, on the other side of the industry, OpenAI is floating the idea that companies should try four-day work weeks to “adapt to the AI era.” Both things are happening simultaneously. We’re simultaneously worried about what employees will do with our data and wondering if we even need employees working five days anymore.

This is what an AI bubble looks like when it hits the real world.

The Infrastructure Is Cracking

Let’s start with the obvious mess: companies have no idea how to present themselves to AI systems. They’ve spent 20 years optimizing for Google’s search algorithm. Now they’re scrambling to reformat their entire web presence because Claude and ChatGPT are crawling everything differently. It’s like watching someone move to a new country and realizing all their professional credentials are useless.

Google’s AI Overviews—those authoritative-looking summaries that appear when you search—pull from “trustworthy sites to Facebook posts,” according to reporting on how accurate they actually are. Facebook posts. Your aunt’s opinion about sourdough starters is now being fed into AI systems that millions of people treat as gospel. The whole information hierarchy we’ve built is collapsing in real time.

Photo by Lucia Barreiros Silva / Pexels

And then there’s the code problem. Companies are drowning in AI-generated code. Not because it’s all good—it’s not—but because it’s there, everywhere, in every system, and nobody’s entirely sure what it does or whether it works. This is what happens when the cost of generating something approaches zero. You get a glut. You get bloat. You get technical debt that’ll take a decade to untangle.

Anthropic thinks their new AI model Mythos is such a security risk that they’re not releasing it publicly. Instead, they’re working with 40 companies in a controlled setting to figure out how bad it could be. That’s not a feature announcement. That’s a fire drill dressed up as a business model.

The Cybersecurity Death Spiral

Here’s the thing about AI and cybersecurity: it cuts both ways, and the offense is winning.

Hackers can now use AI to attack faster. Better reconnaissance. More sophisticated social engineering. Automated vulnerability discovery. The defense is “more AI,” which means companies are buying new systems from Anthropic and OpenAI to try to catch the AI-powered attacks. This is a classic arms race, except the weapons are getting smarter every quarter and most companies’ security teams are still updating spreadsheets by hand.

The Meta data theft wasn’t even fancy. One person with legitimate access just… took the data. That’s not a cybersecurity problem. That’s a people problem. And it suggests that as we layer AI systems into everything, we’re not actually getting better at managing who can access what—we’re just getting faster at moving information around.

Photo by nappy / Pexels

The Work Question Nobody’s Answering Honestly

OpenAI’s four-day work week proposal is either brilliant or dystopian, and I genuinely can’t tell which.

The framing is cute: adapt to the AI era by working less because AI will do more. But here’s what’s actually happening. Companies are using AI to do work that used to require people. Some roles will actually disappear. Others will get restructured into something unrecognizable. And in the meantime, the same people are expected to manage the AI systems, deal with the code overload, try to understand the security implications, and somehow figure out how to present their business to algorithms that barely understand context.

My read is this: four-day weeks aren’t a gift. They’re a pressure release valve. A way to acknowledge that the pace is insane without actually changing the underlying dynamics that made it insane.

What’s Actually Weird About AI Right Now

China’s lobster craze tells you something real about how people relate to AI. Users got an AI assistant and immediately started training it to do what they wanted—raising digital lobsters, basically. In the same month, American teens are harassing chatbots with “funny violence” and confiding about broken hearts to role-playing bots. They’re treating AI like it’s both a toy and a therapist.

That’s not crazy. That’s people filling actual voids. Loneliness. Boredom. The need to feel heard. And they’re finding those things in systems that have no real relationship to them whatsoever.

The US and China are both racing to own AI—different races, different advantages—but the interesting part isn’t who wins. It’s that the winner will own a technology that breaks something fundamental about how we share information, how we work, and how we trust what we see.

Where This Actually Goes

I think we’re six months away from a major AI-driven cybersecurity incident that forces regulation. Not because the technology will be the problem—it’ll be the human element, like the Meta thing, exploited at scale.

I think companies are going to discover that code overload is their biggest technical problem in 2025. Not AI capability. The fact that they have too much of it, nobody understands it, and it all has to be maintained.

I think the four-day week thing catches on, but not for the reasons OpenAI stated. It’ll catch on because companies realize they can’t fill roles, so they consolidate teams and cut hours rather than admitting they’re reducing headcount.

And I think we’re going to see a cultural backlash against AI that doesn’t look like previous tech skepticism. It’ll be weirder. More granular. People will get comfortable with AI in some contexts and absolutely refuse it in others. No universal stance—just “AI for this, never for that.”

The thing that worries me most is the data access problem. We’ve built these systems on the assumption that employees won’t just take stuff. The Meta case proves that assumption is fragile. As AI systems become more valuable, and access more critical, the incentive to steal increases. And the damage from a good theft increases exponentially.

Photo by Markus Spiske / Pexels

What I’m Watching

• Anthropic’s 40-company Mythos trial through Q2 2025. If any of those companies report successful attacks prevented by the model, or conversely, if hackers find ways around it faster than expected, that tells us whether AI defense can keep pace with AI offense. Watch for the first serious breach attempted with AI tools.

• Google’s AI Overview accuracy over the next three months. Specifically: when does a major misinformation incident happen because Google’s system aggregated bad information and presented it as fact? It’s statistical inevitability at this point. The question is timing.

• Enterprise code audit timelines through mid-2025. Companies are going to start scanning their systems for AI-generated code and discovering how much of it exists and how much of it is broken. The ones who report findings first will shape the narrative. The ones who stay quiet will be the ones that really blow up.

• Insider threat incidents at major tech companies. The Meta case is priced in. Watch for whether it triggers systemic changes or just gets treated as an isolated incident. If three more insider threats hit the news in the next six months, you’ll know the industry knows it has a problem it can’t solve with technology.

The AI era isn’t upending the world through capability. It’s upending it through velocity—moving information, code, and systems faster than we can build governance around them. That’s not a technology problem. That’s a human one. And we’re terrible at those.