The Year AI Got Too Dangerous to Control

---

Anthropic built a model so dangerous it won’t let anyone play with it.

That’s the real headline buried under the corporate speak about “access decisions” and “security protocols.” The company developed Mythos, an AI system with such potent hacking capabilities that central banks and intelligence agencies are in emergency response mode. Think about that for a second. We’re past the point of academic debate. We’re at the point where governments are calling each other.

And then—plot twist—someone leaked Claude Code, Anthropic’s AI tool, raising the question nobody wants to answer: if your “too dangerous to release” technology can be leaked, what does “too dangerous” even mean?

This is what happens when the people building the future stop pretending they have a plan.

The Danger That Can’t Be Contained

Let’s be precise about what’s happening. Anthropic created Mythos. It’s powerful. Not theoretically powerful—practically, demonstrably powerful enough that it triggered “emergency responses from central banks and intelligence agencies globally.” Those aren’t words you use lightly. Those are words you use when you’re genuinely worried about financial systems or national security.

So Anthropic did what any responsible company would do: they decided to restrict access. Only approved users. Only vetted purposes. Control the spread.

Then someone leaked Claude Code.

The leaked code doesn’t prove malice. It just proves that airgapping dangerous technology is harder than it looks. And here’s where the contradiction gets sharp: if you can’t keep a model secure among your own employees and partners, how exactly are you gatekeeping it for the broader world? The answer is probably that you can’t, and everyone knows it.

My read is this was predictable. We’ve seen this movie before. The more valuable something is, the more people want it. The more people want it, the more someone will find a way to get it. Anthropic didn’t invent the problem—they’re just running headfirst into it while claiming they can control the outcome.

Photo by Levi Damasceno / Pexels

Meta’s Panopticon Play

Meanwhile, Meta is solving a different problem with maximum creepiness.

The company wants to train better AI models. It needs data. Lots of data. So it’s going to track what its employees do all day—clicks, keystrokes, the works—and feed that into its AI training pipeline. This is technically legal. Morally? That’s another conversation, but Meta’s not waiting for you to have it.

There’s something almost honest about the ruthlessness here. Meta isn’t pretending to be your friend while harvesting your data. It’s literally watching its own employees to build better AI. It’s training humans like they’re datasets.

The weird part is that this probably works. If you want to understand how humans actually work—not how they say they work, but how they really work—watching them type and click is pure signal. It’s behavioural data at scale. It’s the kind of information that’s worth a lot in the AI economy.

But it also tells you something uncomfortable about where we are. We’re past the phase where companies are worried about privacy concerns. They’re past asking for permission. They’re just… doing it. And if employees don’t like it, they can work somewhere else.

When Game Pass Loses Its Bet

Xbox cutting prices on Game Pass while kicking Call of Duty out of day-one access is a smaller story, but it’s important for what it reveals.

Microsoft built Game Pass into a subscription juggernaut by promising everything on day one. That was the whole point. New releases? Included. That drove adoption like nothing else in gaming had managed since, maybe, Netflix in 2007. But subscription models have a math problem: you can’t pay for new AAA games and have a $17-a-month subscription work forever.

So Microsoft’s doing what all subscription services do eventually. They’re pulling back. Call of Duty—one of the biggest franchises on the planet—now arrives “about a year” after release. That’s not a subscription service. That’s a rental shop with a delay.

The price cut is supposed to soften the blow. Cheaper Game Pass, less value per dollar. It’s the kind of move that works in the short term while you’re celebrating the lower churn, then quietly hurts you two years later when everyone realizes the value proposition changed.

I think this is a sign that the streaming wars are done. Not dead—done. The winners are determined. Apple TV+, Disney+, Netflix, HBO Max, and Game Pass all exist in this weird equilibrium where nobody’s winning and everyone’s losing money on originals. The companies are starting to optimize for profitability instead of growth. That means less free stuff, more paywalls, slower adoption.

The Crypto Grift Eats Itself

Justin Sun suing the Trump family’s World Liberty Financial over fraud accusations is basically the moment when even the scammers stop trusting the other scammers.

Sun’s been a major backer of the project. Now he’s claiming fraud. That’s not a disagreement. That’s not a market downturn. That’s “I gave you money for something that was supposed to be real and I don’t think it is.”

The Trump family crypto venture was always going to be a test case for how far brand celebrity could carry a financial product with no real differentiation. Apparently, not far enough. Sun, who’s been in crypto since the Mt. Gox days and knows exactly how these things work, is publicly saying this is fraudulent.

That’s not news because crypto fraud is happening. It’s news because the mask came off in public. The people who profit from crypto hype are now suing each other in court, using the word “fraud,” which is exactly what regulators have been saying all along.

The Unmanned Future Getting Weird

SpaceX is planning an IPO. That’s normal. Companies go public.

But Elon Musk’s shifting the goalposts. The company was supposed to settle Mars. Now it’s supposed to do moonshots that haven’t been defined yet. Also, they’re acquiring Cursor for $60 billion and “emphasizing artificial intelligence.”

None of this is illegal or wrong, exactly. But it’s a sign that when you’re sitting on a $60 billion valuation, the original mission starts to feel quaint. Mars colonization sounds less urgent than “what if we used AI to make something?” The market reward structure is pushing SpaceX away from long-term, defined goals toward open-ended “we’re doing AI stuff” vagueness.

I genuinely don’t know if this makes SpaceX more or less likely to succeed at anything. But it definitely makes it a different company than the one that started twenty years ago.

Photo by nappy / Pexels

What I’m Watching

Anthropic’s next move on Mythos access. The real test isn’t whether they can keep it locked down—they probably can’t indefinitely. It’s whether they’ll publicly admit the gatekeeping failed and start being honest about AI safety being unsolvable by access restriction alone. Watch for any announcement about broader release or a pivot to “we’re focusing on detection instead of prevention.” That shift would tell you everything about whether the company believes its own safety claims.

Meta’s employee AI training rollout and the first lawsuit. Tracking keystrokes will work fine until it doesn’t—which means either an employee class action or a data breach. The specifics don’t matter. What matters is watching whether regulators care. If the EU starts issuing fines by Q2 2025, you’ll know the surveillance capitalism bet is riskier than Meta calculated. If they don’t, you’ll know Meta won.

Call of Duty’s subscription arrival date and Game Pass churn. Microsoft said “about a year.” That’s vague enough to watch for slippage. If it becomes 18 months, if it gets pushed to 2026, Game Pass is in trouble. Also track Game Pass subscriber numbers in Q1 and Q2 2025—if the price cut doesn’t arrest churn, the whole model’s unraveling.

Whether Justin Sun’s lawsuit actually goes to trial. Crypto lawsuits usually settle quietly. A public trial would be the most honest thing to happen in crypto finance in a decade. The discovery phase alone would be worth watching.