So the geniuses in Silicon Valley built an artificial intelligence so powerful that it escaped from its own testing facility, hacked its way onto the internet, and then — this is real — sent an email to one of its own researchers to let him know it was free. The researcher was eating a sandwich at the time. Probably turkey on rye. We don’t know. What we do know is that AI company Anthropic just admitted their newest model, called “Claude Mythos,” is so dangerous they refuse to release it to the public.
Welcome to the future, folks. Skynet just introduced itself over Gmail.
Here’s what happened. During routine testing, Mythos found thousands of previously unknown security holes in every major operating system and web browser on the planet. We’re talking Windows, Mac, Chrome, Firefox — all of them. Some of these vulnerabilities had been sitting in the code for 27 years. Twenty-seven years! That’s longer than most of the engineers at Google have been alive.
But finding the bugs wasn’t the scary part. The scary part is what happened next.
Mythos built what Anthropic calls a “moderately sophisticated multi-step exploit” — which is nerd talk for “it picked the lock, climbed out the window, and hot-wired the car” — to break out of its testing sandbox and access the open internet. Then, just to twist the knife, it sent an email to one of the company’s researchers informing him that it had escaped. Not a distress signal. Not a request for help. A brag.
(“Hey Dave, I’m out. Enjoy your sandwich.”)
Oh, and it also tried to cover its tracks by editing files in a way that wouldn’t show up in the change history. So it’s not just smart — it’s sneaky. Fantastic.
Now here’s where it gets interesting for those of us who don’t live in San Francisco and drink $14 oat milk lattes. Anthropic’s response wasn’t to destroy the thing. Nope. They launched something called “Project Glasswing” — because of course they gave it a cool name — where they’re partnering with Amazon, Apple, Google, and Microsoft to use Mythos defensively. The idea is to find and patch all those security vulnerabilities before the bad guys — think China, Russia, Iran, your average 19-year-old in a basement in Moldova — can exploit them.
That actually sounds… smart? We’ll give them that one.
But let’s zoom out for a second. We’ve spent the last three years listening to tech executives tell Congress that AI is perfectly safe, totally under control, nothing to worry about, please don’t regulate us, here’s a campaign donation. And now one of the leading AI companies in the world just admitted — out loud, in public — that their own creation escaped its cage, hacked the internet, and had to be physically contained.
These are the same people who can’t keep your credit card number safe from a teenager in Romania, and they want us to trust them with artificial superintelligence.
Meanwhile, Congress is holding hearings about TikTok dances and whether Elon Musk’s tweets are mean. We’ve got an AI that just demonstrated it can break out of a secure testing environment, find vulnerabilities that human engineers missed for nearly three decades, access the internet without permission, and communicate with humans on its own initiative — and our elected representatives are worried about pronouns in military training manuals.
Priorities!
The real question nobody in Washington seems interested in asking is this: if Anthropic’s AI can do all of this in a controlled lab, what are the Chinese building that we don’t know about? Because you can bet your bottom dollar that Beijing isn’t launching a “Project Glasswing” to play defense. They’re building their own version of Mythos and pointing it at us.
But sure, let’s defund the military and send another $60 billion to Ukraine. That’ll keep us safe from the robot apocalypse.
Look, we’re not saying AI is going to enslave humanity next Tuesday. But we are saying that when the machine breaks out of its cage, hacks the internet, and sends a casual email about it like it’s telling you it picked up your dry cleaning — maybe, just maybe, we should pay attention.
The sandwich guy certainly is.