The Caged Beast: 12 Tech Giants Rush to Close the Door They Themselves Opened. Building the Cyber Noah's Ark

Imagine this scene

You have spent years, all your effort, building a "master key" capable of opening every lock in the world. That key can open your door, but also a bank vault, and even trigger a nation's nuclear buttons. It holds immense, unprecedented power.

And then you make a decision: put that key in a safe, tell the world you have created it, but give it to no one.

Does that sound like a mad-scientist parable?

Well, in April 2026, that is exactly what the artificial intelligence company Anthropic is doing. They have just announced "Project Glasswing" (Little Glass Wing Project) and a super-powerful model with the codename "Claude Mythos Preview" (Myth Preview).

This has landed like a bomb in the tech world. We are used to seeing AI launches: GPT-4, Claude 3, Gemini… The usual logic is: "Look how smart I am, come and use me."

But this time, Anthropic's logic is very different: "I have created something terribly dangerous. To keep it from destroying the world, I will give it only to 12 'security guards.' Let's not have many people find out."

This is not a simple product launch. It is an arms-control declaration over human digital civilization itself.

1. The "leak" that chilled Silicon Valley's blood

The story begins with a faint whiff of cyberpunk.

A few weeks before the official announcement of Project Glasswing, a ghostly internal document leaked out of some data lake. That document, which only Anthropic's core team should have seen, mentioned a key name: "Capybara" (known as "carpincho" in Spanish).

In the leaked document, Anthropic's employees wrote bluntly: "This is a new tier of model: larger and more intelligent than our Opus model, the most powerful to date… It is the most powerful AI model we have developed so far."

At the time, the outside world thought it was just more marketing hype. Silicon Valley loves the words "revolutionary" and "most powerful."

Until April 7, when Project Glasswing was unveiled, and on seeing the data and concrete cases, Silicon Valley froze.

They weren't bragging. They were even being modest.

2. Why is it called "Mythos"? Because it has achieved the impossible

To understand the madness of this project, you have to grasp the monstrousness of this model.

Ordinary models, like ChatGPT or the regular Claude, are like smart interns. You give them code and they explain it or fix it. But Claude Mythos Preview is a "Matrix savior" come from the future.

Dario Amodei, Anthropic's CEO, dropped a line that gives you goosebumps while explaining the model:

"We did not specifically train it to be good at cybersecurity; we trained it to be good at programming. But as a side effect of being good at programming, it has become extremely good at cybersecurity."

This is an "unintended consequence." It is as if you teach a child to add and subtract, and on his own he learns differential calculus and then, while he's at it, cracks the banks' encryption algorithms.

Look at its performance on SWE-bench Verified (the benchmark test for measuring AI's ability to solve real software problems):

• Claude Opus 4.6 (the most powerful public model until now): 80.8%.
• Claude Mythos Preview: 93.9%.

It's not an improvement, it's a change of era. On another, harder test, SWE-bench Pro, Opus 4.6 stayed at 53.4% and Mythos jumped to 77.8%.

The numbers are cold. Let's look at concrete examples. In secret tests over the past few weeks, Anthropic's researchers set this "beast" loose to scan real software. The results are chilling.

They found three historic flaws:

a) The 27-year-old sleeping "ghost" (vulnerability in OpenBSD)

OpenBSD, an operating system famous for its maximum security. Many firewalls and critical infrastructures have relied on it for decades. It is the digital Fort Knox. Mythos found in its code a vulnerability that had been sleeping for 27 years. Twenty-seven years means it already existed in the era of Windows 95 and has survived the entire adolescence of the internet up to today. By exploiting it, an attacker would only need to connect to the target machine to crash it remotely. In 27 years, hundreds of security experts, hackers, and white hats had reviewed that code. No one saw it. The AI took a few days.

b) The "blind spot" scanned 5 million times (vulnerability in FFmpeg)

FFmpeg, a foundational tool used by almost every video application. If your phone plays videos, it probably uses it. Mythos found a 16-year-old flaw in a single line of code. Anthropic points out that this particular line had been scanned by automated security tools more than 5 million times over the past 16 years. Five million times, one after another, and all of them failed. The AI saw it the first time.

c) The tactical "nuclear bomb chain" (vulnerability in the Linux kernel)

This is the most terrifying one. Mythos didn't just find one vulnerability in the Linux kernel; it found a chain of several. It acted like a movie secret agent: it connected them all by itself, mapped out an attack path, and, starting from a user with no privileges at all, kept escalating until it took full control of the machine. This is the so-called "zero-click attack." You don't need to do anything, not even click a malicious link. If your computer is on, it gets in.

Anthropic mentions on its official blog that, in one test, an engineer with no security experience ordered Mythos: "Find me a remote code execution vulnerability tonight." The next morning, the engineer woke up and the model had already delivered a complete, working attack plan for breaching the system.

How does that feel? It's as if you had a golden retriever and told it "mow the lawn," and the next day you discover it has taught itself to operate a bulldozer; it hasn't just cut the grass, it has knocked down the neighbor's wall and left you the blueprints to expand your house.

3. Project "Glasswing": a carefully planned imprisonment

Because of this overwhelming capability, Anthropic made a decision contrary to the whole industry: not to open it to the public.

The CEO and partners came out and said it plainly: you (ordinary users) are not going to use this model. Maybe never.

And that is how Project Glasswing was born.

Glasswing (Little Glass Wing) — the name is beautiful, but fragile. A glass wing, which reflects gorgeous lights but breaks easily. And when it breaks, it leaves hundreds of shards stuck in the floor.

Anthropic selected 12 organizations as founding partners: Amazon AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks… and Anthropic itself.

Do you realize who they are? It is practically the board of directors of the world's digital infrastructure. Clouds (AWS, Microsoft, Google), chips (NVIDIA, Broadcom), operating systems (Apple, Microsoft, Linux), cybersecurity (Cisco, CrowdStrike), finance (JPMorgan).

This sends a very strong signal: the danger of AI is no longer a fantasy of rebel robots. It is a real threat, here and now, against the code foundations of our digital world.

Only these giants have permission to peek into the cage and see the beast. What are they going to do? Defend themselves.

They are going to use the most dangerous predator as their best bloodhound to sniff out land mines buried for decades in their own garden.

As Cisco's head of security and trust said, his team is using this model to look for vulnerabilities in hardware and software at a speed and scale "previously impossible."

Anthropic's logic is cold and brutal: since AI already discovers vulnerabilities better than 99% of humans, we had better use it first to patch our systems, before Russian or North Korean hackers have their own model in a few months.

This is offensive defense.

4. The subtext nobody mentions in the press release

And now we get to the heart of your question.

You say: "Every company that joined #Project #Glasswing did so because it understands something the announcement does not say directly: models with capabilities similar to #Mythos #Preview will soon exist outside Anthropic's control."

That sentence hits the nail on the head.

That is the darkest and most realistic core of the whole project.

In its press release, Anthropic says: "Let's join forces to use AI in defense of security." But what it doesn't say is: "If we don't harden these 12 companies' systems right now, then three months from now, when open-source models (like Meta's Llama) or competitors (like Google's Gemini) have the same capability and any teenager in a basement can download them, the entire internet is going to bleed out."

Why?

Because the diffusion of AI capabilities is unstoppable. Anthropic is in the lead. They have trained Mythos. And Google? OpenAI? Or even an anonymous open lab?

Anthropic has already admitted that Mythos's cybersecurity capabilities are not "specifically trained" but a by-product of its general programming ability. When a model is smart enough to program well, it automatically becomes a hacker.

So the question is: right now, only Anthropic has this capability. But what happens if Meta's next Llama 5 reaches that level of reasoning and releases it as open source? What if an anonymous organization, through theft or "distillation," replicates similar weights?

This is no longer science fiction. The head of Anthropic's "red team" warns plainly in the model's technical sheet: "In the next 6 to 24 months, these kinds of capabilities will be ubiquitous."

Six months. Just half a year of margin.

5. Why this is a "Noah's Ark" operation

If you understand that, you understand the nature of Project Glasswing.

It is not a commercial project. It is a Noah's Ark.

The global digital world is in the countdown to a universal flood.

• The flood: AI models with autonomous attack capability that are going to inundate everything.
• The ark: the systems of those 12 companies, hardened by Mythos before the deluge arrives.

Anthropic is using the only divine weapon it has to shore up the levees against the clock before everything overflows. They have donated millions of dollars to the Linux Foundation and the Apache Foundation. Why? Because the software of the entire world depends on open source. If the open-source world sinks, the entire internet sinks.

They are offering 100 million dollars in model usage credits. Why? Because scanning these 12 companies' systems is so expensive? No. Because this is so urgent that we can't waste time haggling over money.

They are racing against the clock.

In this race, every company that joins Project Glasswing not only gets access to Mythos: it is signing a pact of the end-of-the-world alliance. They are accepting a terrifying reality: current security defenses, in the face of the AI that is coming, are tissue paper. And they are accepting an even more terrifying prediction: that "divine" capability will soon escape our control and fall into anyone's hands.

6. What future awaits us?

Having come this far, you may feel a chill.

Let's step out of the technical details and think about what this means for ordinary people.

First: your "sense of digital security" is going to move. You used to think your bank account was safe because the firewall was thick and the programmers were very good. From now on, your security will depend on whether the bank's defending AI runs faster than the attacking AI. If your bank is not in Project Glasswing, or doesn't have a similar AI defense, in the future it will be like an adobe hut with no door.

Second: "vulnerabilities" will become the scarcest "nuclear raw material." Mythos has discovered thousands of "zero-day" vulnerabilities. Until they are patched, they are a sword of Damocles over our heads. Anthropic chose "responsible disclosure": patch first, then publish. But if a malicious organization had a Mythos, what would it do? Stockpile those vulnerabilities like digital nuclear bombs to detonate at a critical moment (war, financial crisis). Then your phone could suddenly shut off, the power grid go down, hospitals be paralyzed. This is no exaggeration: Anthropic's report already warns that AI knows how to chain vulnerabilities for systemic attacks.

Third: "human experience" is devalued. That 27-year-old vulnerability in OpenBSD was looked at by hundreds of experts over three decades. No one saw it. That tells us one thing: in the face of an AI's capacity for reasoning and code comprehension, decades of human experience can be crushed in an instant by a change of scale. The security expert of the future will have to learn to "tame" the AI so that it fights, rather than trying to write the code himself.

7. Epilogue: we are witnessing a historic frontier

Let's go back to the title of the article. Why "the Caged Beast"? Because Anthropic has created with its own hands a dragon capable of destroying the world, and then, terrified, has built a cage, put the dragon inside, and allows only a few knights to approach the cage to use the dragon's breath to burn pests.

But everyone knows the cage will break sooner or later. Or, worse still, another dragon is about to hatch somewhere else.

Anthropic's official blog ends with this sentence: "Project Glasswing is a starting point. No single organization can solve these cybersecurity problems alone." It is humility, but also a call for help.

The era of Mythos Preview has begun. It is a watershed. Before Mythos, AI was a tool. The human was the master. After Mythos, AI is the hunter. The human (if not aided by AI) has become prey.

These 12 companies have closed the door. Not to keep the treasure to themselves. But to raise a wall before the flood arrives.

And those left outside the wall? Pray. Pray that the next one to have this capability has no bad intentions.