The Year AI Got Too Dangerous to Ship

---

Anthropic built something so dangerous they won’t let anyone use it.

That’s the actual state of AI development right now. Mythos, their latest model, has triggered emergency responses from central banks and intelligence agencies globally. The company decided to keep it locked down because of its hacking capabilities. Not theoretical hacking. Actual, demonstrated ability to break into systems.

This is the moment the industry stopped pretending.

For a decade, we’ve watched tech move at velocity and ask forgiveness later. Move fast, break things, settle the lawsuits in 2028. But something’s shifted in the last few months. The scale of what these systems can do has finally caught up with the scale of what the companies are willing to release, and the gap is terrifying.

Photo by Solvita Vita / Pexels

When Keeping It Secret Is the Product

Here’s what’s wild about the Mythos situation: Anthropic isn’t hiding it because the model might generate a mean email. They’re hiding it because it can actively infiltrate systems. Central banks. Intelligence agencies. These aren’t entities that scare easily. If they’re issuing emergency responses, the threat isn’t hypothetical.

The leak of Claude code raises a different angle. Someone got their hands on Anthropic’s work, and now there’s a live debate about whether copyright even matters in an AI era. The code’s out there. It’s faster than ever to reproduce creative work with these tools. Does the legal framework we built for books and songs even apply?

My read: we’re about to find out that it doesn’t, and that’ll be a much bigger problem than anyone’s admitting.

The UK Biobank hack is the physical-world equivalent. 500,000 people’s medical data got lifted and listed for sale in China. The bank claims no personally identifiable information was compromised, but that’s the kind of statement that ages like milk. In three months, we’ll learn what “no PII” actually meant—probably something like “we didn’t include names, but the health records alone would identify half these people.”

The Surveillance You Actually Agreed To

Meta’s decision to track workers’ clicks and keystrokes for AI training is something else entirely. This isn’t a breach. It’s not unauthorized. It’s policy.

Think about that. A company is openly taking the data it can see about how humans work and feeding it directly into AI models. Not anonymized. Not aggregated into behavioral patterns. The actual keystroke sequences, the actual decision points, the actual timing of how a person moves through a workday.

This is Meta testing out infrastructure. If it works on employees—people who can’t really opt out without losing their job—it’s a proof of concept for everything else. Every other company watching this is taking notes.

I think this is where the industry’s patience with regulation finally breaks. Not because it’s illegal. Because it’s too honest. When companies hide surveillance behind Terms of Service, regulators get distracted arguing about consent. When a company just straight-up says “we’re watching everything you do,” the conversation becomes uncomfortable.

Photo by nappy / Pexels

The IPO That Admits the Problem

SpaceX is going public. That’s the headline. But the detail is more interesting: Musk’s goals have shifted away from Mars and toward moonshots Anthropic-adjacent in scope.

The company’s been emphasizing AI more heavily. Then it struck a deal with Cursor for $60 billion. That’s not a small acquisition. That’s not a side project. That’s SpaceX pivoting toward AI infrastructure while simultaneously preparing to be a public company.

Here’s my uncertainty: I can’t quite figure out if this is visionary or desperate. Is Musk seeing the future and positioning SpaceX as essential infrastructure? Or is he recognizing that Mars is a 30-year problem and investors want ROI in 5-7 years?

Probably both. That’s what makes it interesting.

The broader pattern though—companies going public while their risk profile is accelerating rather than stabilizing—feels dangerous. Investors get quarterly earnings calls. They don’t get emergency briefings from central banks about hacking capabilities.

The Copyright Question That’ll Reshape Everything

Here’s the thing about the leaked Claude code and the copyright questions swirling around it: this isn’t abstract anymore.

AI tools are making it faster than ever to reproduce creative work. A songwriter used to take weeks to arrange a piece. Now it takes minutes. A designer used to need references and months of skill-building. Now they feed a prompt and get something useful in seconds.

The legal question is dead. Copyright doesn’t matter for the infrastructure. What matters is what society decides to do about it. And I think we’re about to learn that “what society decides” really means “what the people with leverage decide.”

Artists don’t have leverage. Programmers have a little more. Companies have all of it.

The phishing angle in all this—hackers spoofing Paperless Post and Evite to break into systems—shows the other layer of the risk environment. While we’re debating whether AI should be regulated, people are still getting compromised through invitation-based social engineering. The new threats aren’t replacing the old ones. They’re layering on top.

What Actually Matters Now

Microsoft ending day-one Game Pass access for new Call of Duty games is a small move with a big signal. The company’s cutting prices but delaying AAA releases by a year. Translation: even Microsoft is admitting that the subscription model has limits. You can’t give people everything immediately and charge them eight dollars a month.

It’s a sign of market maturity. Services reach a ceiling. When they do, companies start protecting the premium product.

Peter Molyneux saying Masters of Albion will be his last game—that’s a different kind of signal. The creator of Fable and Black & White is stepping back. He’s 61. He’s saying the industry is changing in ways that don’t excite him anymore. When someone who built some of the most imaginative games of the 2000s says he’s done, it’s worth listening to why.

Photo by Markus Spiske / Pexels

The pattern across all these stories is the same: systems are getting too powerful to move fast with them. Too powerful to release without thinking. Too powerful to train without oversight. Too powerful to ignore.

Anthropic’s locked-down Mythos isn’t a setback. It’s a maturation moment. It’s what happens when the thing you built might actually break the world if you’re wrong about containment.

My prediction: by Q4 2025, we’ll see the first real regulatory action that actually sticks because it’ll target something this obvious—either the surveillance infrastructure (Meta’s click-tracking will become the flashpoint) or the security risks (Mythos-level models will have to be registered like military hardware). The companies won’t like it. The public won’t understand it. But it’s coming because the status quo is unsustainable.

And when it arrives, it’ll feel like it took forever and happened overnight.

What I’m Watching

• Meta’s employee surveillance rollout timeline. Watch for when this expands beyond internal testing. If other companies start copying it within six months, we’ve got a new normal. If regulatory friction hits within three months, we’ve got the catalyst for broader policy.

• The first Mythos-level model release. Someone else will build something equally dangerous. Will they release it like Anthropic did (locked down), or will there be competitive pressure to open it up? That decision point—and how regulators respond—is the actual inflection.

• SpaceX’s IPO prospectus language around AI. Specifically: how much of the company’s future value they attribute to Cursor vs. rockets. If it’s more than 30%, Musk’s officially betting on a different future than he used to.

• Copyright litigation around AI training. The first major lawsuit will probably land in Q2-Q3 2025. Watch whether courts rule on the training itself or just the distribution. The technical scope of that ruling determines whether AI development stays in the US or moves somewhere with fewer restrictions.