Imagine a manmade intelligence so powerful that it manages to discover security vulnerabilities that have been hidden for decades in the most critical systems on the planet, in a matter of hours and without human intervention. That’s exactly what he did Claude Mythosthe new Anthropic model that is shaking the entire technology industry, and that the company itself admits that cannot be released to the public because of the risks it represents.
The story began almost by accident. In late March 2026, Anthropic inadvertently leaked information about the existence of this model through an internal data leak. What came to light left more than one with their mouths open: Mythos is not an incremental update of Claude.
It is, according to the company itself, “a level change” in performancea completely new category above its Opus models, with capabilities that far exceed anything on the market in areas such as academic reasoning, advanced programming and, most worryingly, cybersecurity.
Claude Mythos: the AI that found thousands of Zero Day vulnerabilities alone
This is where things get serious. Anthropic’s security team released a detailed technical assessment of Mythos Preview’s capabilities after about a month of internal testing, and the results are, frankly, chilling.
Mythos found thousands of Zero Day vulnerabilities —that is, security flaws that no one had discovered before—in critical map infrastructures of the entire world. And we’re not talking about minor bugs: the model identified a denial of service vulnerability in OpenBSD that carried 27 years without detectiona flaw in FFmpeg’s H.264 codec introduced in 2003 and exposed in 2010 that had gone unnoticed by every fuzzer and human researcher who had reviewed that code, and a FreeBSD remote code execution bug from 17 years old which allowed root access without authentication. He did all of this autonomously, without anyone telling him step by step how to do it.
To measure the difference with previous models, the researchers did a direct test: they put Mythos Preview and its predecessor, Opus 4.6, against the same vulnerabilities in the Firefox JavaScript engine 147. Opus 4.6 managed to generate functional exploits barely twice in hundreds of attempts. Mythos Preview did it 181 times. It is not a slack improvement. It’s a jump from another dimension.
The most disturbing thing about all this is that Anthropic admits that he did not train Mythos specifically to have these abilities. They emerged emergently as a consequence of general improvements in reasoning, code, and autonomy. In other words: the same skills that make you extraordinarily good at patching vulnerabilities make you equally good at exploiting them.
Why Anthropic can’t release Mythos to the public yet
The problem is obvious: if an AI can find and exploit Zero Day vulnerabilities on all major operating systems and browsers standalone, in the wrong hands it would be a digital weapon of mass destruction. Anthropic knows this and says it bluntly.
The company noted in its leaked internal communications that Mythos is “currently far ahead of any other AI model in cyber capabilities”and that “heralds a next wave of models that can exploit vulnerabilities in ways that far exceed the efforts of defenders.” In other words: what Mythos can do today, other less responsible models could do tomorrow.
And the risk is not theoretical. Anthropic has previously documented that malicious actors have already attempted to use their existing models for coordinated attacks against dozens of organizations simultaneously, including financial institutions and government agencies. Now imagine that with Mythos.
Therefore, the decision was clear: Mythos is not released to the public current. At least, not yet. But they couldn’t just keep it in a drawer knowing that the map world ecosystem is full of vulnerabilities that the model already knows how to detect. They had to do something with that knowledge, and do it quickly.
Project Glasswing: the technology industry’s response to the Mythos challenge
That’s how it was born Project Glasswingand the scale of the initiative says a lot about the seriousness of the problem.
Anthropic contacted the industry giants and on April 7, 2026 officially launched this collaboration that brings together, among others, Amazon Web Products and companies, Apple, Google, Microsoft, NVIDIA, Cisco, CrowdStrike, Broadcom, JPMorganChase, the Linux Foundation and Palo Alto Networks. Additionally, more than 40 additional organizations involved in critical map maintenance also received access.
The idea is simple but powerful: give defenders a head start before models with similar capabilities are available to anyone. Project partners can use Claude Mythos Preview to scan their infrastructure, detect weaknesses, and improve their security practices. Anthropic, for its part, committed to investing up to $100 million in usage credits for participants, plus $4 million in direct donations to open source security organizations.
The partners’ statements are revealing. Microsoft confirmed that when they tested Mythos against their open source security benchmark CTI-REALM, the model showed “substantial improvements” compared to previous generations. From CrowdStrike they were even more direct: “The window between a vulnerability being discovered and an adversary exploiting it has collapsed: what used to take months now happens in minutes with AI.” And from the Linux Foundation they warned that this is not just an opportunity; It is also a sign that attackers will soon have access to similar capabilities.
The launch of Project Glasswing marks a historic moment for cybersecurity: for the first time, frontier-level AI is deployed not as a commercial product, but as collective shield of world digital infrastructure. The question we are all asking now is how long it will be before the capabilities of Mythos are available to those who do not have such noble intentions.
Keep reading:
• Anthropic, the AI company that confronted the Pentagon in the US and why this concerns us all
• Anthropic mistakenly leaked Claude Code’s source code and the damage could be enormous
• Claude’s Talents: the feature that completely changes how you use AI
