Anthropic Built an AI So Dangerous They're Refusing to Release It. What They Found Will Terrify You.

Sleep well last night? Enjoy it. Because your computer, your phone, your bank they've all been vulnerable for years. Decades, even. And now we know exactly how bad it really is.

Anthropic just dropped a bombshell that the tech world is scrambling to process. They've built a new AI model called Claude Mythos Preview, and it's so terrifyingly powerful at hacking that they won't let the public anywhere near it. Instead, they panicked, called up Apple, Google, Microsoft, Amazon, NVIDIA, and a half-dozen other tech giants, and formed an emergency $100 million coalition called Project Glasswing to quietly patch the internet before someone else builds something like this.

Read that again. A hundred million dollars. To fix what one AI found overnight.

Here's what Mythos did that has Silicon Valley losing sleep:

It found thousands of high-severity vulnerabilities in every major operating system and every major web browser. Windows. macOS. Linux. Chrome. Safari. Firefox. All of them. Riddled with holes that have been sitting there, some for decades, waiting for something smart enough to find them.

The numbers are absolutely insane. For about fifty dollars in compute costs, Mythos found a 27-year-old bug hiding in OpenBSD widely considered one of the most security-hardened operating systems ever built. Security researchers have been auditing that codebase for nearly three decades. An AI found what they missed for the price of a pizza dinner.

It gets worse. For under $1,000, it didn't just find a vulnerability in FreeBSD it built a fully working exploit. Remote code execution. Unauthenticated root access. From anywhere on the internet. That means anyone, anywhere, could theoretically take complete control of an affected system without needing a password.

And for under $2,000? It chained together multiple Linux kernel vulnerabilities into a complete privilege escalation exploit. The kind of attack chain that elite government-funded hackers spend weeks crafting done by an AI before the engineers who launched it finished their morning coffee.

Here's the part that should genuinely scare you: the Anthropic engineers who ran these tests had no formal security training. They just asked Mythos to find exploits, went to bed, and woke up to working attack code. That's it. That's all it took.

Now imagine what happens when this kind of capability isn't in the hands of a company trying to do the right thing. Imagine a rogue nation-state. A ransomware gang. A teenager with a grudge and a credit card. Anthropic knows this is coming. That's why they're racing against the clock.

Project Glasswing has already extended access to over 40 additional organizations that maintain critical software infrastructure. Anthropic is throwing $100 million in compute credits and $4 million in direct donations at open-source security projects. They're essentially saying: the fire is already lit, we're just trying to build firebreaks before it spreads.

The uncomfortable truth? This is just the preview model. The next generation will be even better at this. And eventually, someone who doesn't care about responsible disclosure will build one.

The clock is ticking. And your software was never as safe as you thought.