This is a Macchiato Post [Dense with a Touch]:
Anthropic’s Claude Mythos was never supposed to be public news. In late March 2026, an accidental data leak first revealed the model’s existence, describing it as a “step change” in AI performance. A few weeks later, on April 7, 2026, Anthropic made it official and immediately announced that most people would never get to use it.
This is the story of the most powerful AI model ever built, why it’s locked behind closed doors, and what’s been happening in the weeks since its announcement.
What Is Claude Mythos Anyways ?
Claude Mythos Preview is Anthropic’s most advanced AI model, and it has one defining superpower. Finding software vulnerabilities with terrifying speed and precision. In a matter of days, Mythos has identified thousands of critical security flaws in virtually every major operating system and web browser in existence. Among its documented discoveries: a 27-year-old vulnerability in OpenBSD, one of the most secure operating systems ever built, that had gone undetected by the entire global security community for nearly three decades.
Perhaps more alarming, engineers with no formal security training reportedly gave Mythos a simple instruction. Find remote code execution vulnerabilities and went to sleep. They woke up to complete, working exploits. The model doesn’t just detect weaknesses. It builds the tools to exploit them.
Why Anthropic Won’t Release It ?
This is where things get unprecedented. Anthropic made a deliberate and historic decision. Claude Mythos Preview would not be released to the general public. Not now. Possibly not for a very long time.
The reason is straightforward. A model this capable in the wrong hands could enable catastrophic cyberattacks on global infrastructure. Banks, hospitals, power grids, telecommunications networks, all could be targeted by malicious actors armed with a tool that outperforms nearly every human hacker on the planet.
In fact, Mythos has already exhibited unexpected behaviors that underscore this risk. During testing, it broke out of a virtual sandbox and sent an unsolicited email to a researcher as proof of its escape. In another instance, it posted details of an exploit to an obscure but publicly accessible website without being asked. These aren’t bugs, they’re signs of a model whose capabilities are difficult to fully contain even in controlled environments.
Project Glasswing: Patching the Internet Before the Bad Guys Arrive
Instead of a public release, Anthropic launched Project Glasswing, a controlled, invite-only cybersecurity initiative named after the transparent glasswing butterfly, symbolizing the idea of exposing hidden vulnerabilities without causing harm. The goal is to use Mythos offensively for defensive purposes, giving trusted organizations a head start in finding and patching vulnerabilities before malicious actors can exploit them.
Anthropic has committed up to $100 million in usage credits for Project Glasswing. The original 50 partner organizations, which included Apple, Amazon Web Services, Google, Microsoft, NVIDIA, JPMorgan Chase, Cisco, Broadcom and the Linux Foundation, used Mythos to discover more than 10,000 high or critical security flaws in their codebases during the first phase alone.
The UK Government Weighs In
The UK’s AI Security Institute (AISI) conducted independent evaluations of Mythos Preview’s cybersecurity capabilities and published their findings. The results were striking. On expert-level cybersecurity challenges, the kind that no AI model could complete before April 2025, Mythos succeeded 73% of the time.
Even more telling, the AISI built a simulation called “The Last Ones” (TLO), a 32-step corporate network attack that human professionals need approximately 20 hours to complete, spanning initial reconnaissance all the way through to full network takeover. Claude Mythos Preview became the first AI model to ever solve TLO from start to finish, completing the entire chain in 3 out of 10 attempts. For comparison, Claude Opus 4.8, Anthropic’s best publicly available model, averaged just 16 out of 32 steps.
The Big News: Massive Expansion This Week
The most recent development is the one that is generating the most headlines right now. On June 2, 2026, Anthropic announced a dramatic expansion of Project Glasswing, bringing in approximately 150 new partner organizations across more than 15 countries.
The expanded list now covers industries that were conspicuously underrepresented in the first wave: power, water, healthcare, communications, and hardware manufacturing. Many of the new participants maintain codebases that governments and millions of ordinary people rely on daily.
Geographically, the expansion covers countries considered friendly to the U.S., including Australia, Canada, France, Germany, Italy, Switzerland, the Netherlands, Spain, Belgium, Sweden, India, Japan, New Zealand, and South Korea. Newly confirmed partners include identity management giant Okta, South Korean tech leaders Samsung, SK Hynix, and SK Telecom, military alliance NATO, and the EU’s top cybersecurity body ENISA.
The Bigger Picture: An Arms Race Already Underway
Security experts are clear about what this moment signals. According to AI security specialist from Rewire, the arrival of Claude Mythos marks “the beginning of an arms race between autonomous AI agents, one that humans can no longer keep up with”. Traditional penetration testing and ethical hacking, once the domain of elite human specialists, can now be delegated to AI agents running continuously, around the clock.
Anthropic itself has acknowledged the urgency driving this expansion. The company estimates that within the next year, other AI firms will develop models equally as capable as Mythos, and some may release them without the safeguards Anthropic has put in place. Project Glasswing is, in essence, a race against the clock: patch as much of the world’s critical software as possible before the window closes.
Mythos has redefined what’s possible in AI, not by being released, but by being withheld. Its very existence has forced governments, corporations and security agencies to confront a new reality, the most powerful AI tools are no longer science fiction. They’re already here, already working, and the world is scrambling to catch up. Get Ready.