Artificial intelligence just crossed a line that most experts knew was coming — but few expected to arrive this fast. Anthropic has officially unveiled Claude Mythos Preview, its most powerful AI model to date, and the company is so concerned about its capabilities that it has decided against a public release entirely. Instead, Anthropic launched Project Glasswing, a tightly controlled cybersecurity initiative that grants a select group of the world’s biggest tech companies early access to the model—specifically to find and fix critical vulnerabilities before bad actors can weaponize the same technology.

I’ve been following the AI race closely for a while now, and honestly, this is one of the most consequential announcements in the space so far this year. It’s not just another model drop. It’s a signal that frontier AI has entered territory that requires an entirely new playbook—one the industry is still scrambling to write.

What Is Claude Mythos — and Why Did It Leak First?

The story of Claude Mythos doesn’t begin with a press release. It begins with a data leak. In late March 2026, Fortune discovered Anthropic’s draft blog post sitting in an unsecured, publicly searchable data cache—a result of human error within the company’s content management system. That post described Mythos as “by far the most powerful AI model” the company had ever developed and warned it poses unprecedented cybersecurity risks. Naturally, the tech world took notice immediately.

What most articles missed is just how severe the market reaction was. Shares in CrowdStrike, Palo Alto Networks, Zscaler, SentinelOne, Okta, Netskope, and Tenable all fell between 5% and 11% following the leak, as investors worried the model could undermine demand for traditional security products. That kind of selloff doesn’t happen over hype—it happens when something real and serious is sitting behind the curtain, threatening to change the industry’s entire value proposition.

Two weeks later, Anthropic made the official announcement on April 7, 2026. Rather than a standard public launch, the company introduced Project Glasswing—a new cybersecurity coalition that frames Claude Mythos not as a product release but as a global defensive mission.

Project Glasswing: The Biggest Tech Names United for Cyber Defense

The Project Glasswing initiative brings together Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks—all working under a single defensive cybersecurity umbrella. That is an extraordinary coalition. Essentially a who’s who of global digital infrastructure, assembled to fight an invisible war inside the world’s most critical software systems.

Anthropic is providing up to $100 million in usage credits to the companies testing Mythos Preview, and an additional $4 million to open-source security organizations, including OpenSSF, Alpha-Omega, and the Apache Software Foundation. That level of financial commitment tells you everything about how seriously this is being taken. This isn’t a pilot program—it’s a full mobilization of the technology industry’s top tier.

What I find interesting here is how the name “Project Glasswing” was chosen. Employees at Anthropic selected it as a metaphor likening a transparent butterfly to software vulnerabilities, which are “relatively invisible.” It’s a subtle but fitting analogy—these bugs have been hiding in plain sight inside systems that billions of people depend on every single day.

Claude Mythos Has Already Found Thousands of Zero-Day Vulnerabilities

Here’s where the numbers get genuinely alarming. In just a matter of weeks, Anthropic says Claude, the Mythos Preview, has identified thousands of zero-day vulnerabilities—many of them critical and difficult to detect—including bugs in every major operating system and web browser currently in use. To put that in perspective, Claude Opus 4.6, the last model Anthropic released to the public, found approximately 500 zero-days in open-source software. That’s a fraction of Mythos Preview’s output. This isn’t an incremental improvement—that is a completely different class of capability.

The buried stat that most outlets glossed over? Mythos Preview successfully reproduced vulnerabilities and created proof-of-concept exploits on the first attempt in 83.1% of cases. An 83% first-attempt success rate on working exploits is the kind of number that keeps security researchers awake at night. The oldest vulnerability found? A 27-year-old bug in OpenBSD—an open-source operating system widely regarded for its strong security record—had survived decades of human review and millions of automated security tests. Mythos found it in weeks.

After looking into this more closely, I can tell you that the gap between Claude Mythos and every other publicly available model is wider than almost any previous generation jump in AI capability. One test involved finding and chaining together multiple flaws in the Linux kernel—the operating system powering the majority of the world’s servers—in a way that would allow a hacker to take complete control of any machine running it. That’s not a theoretical risk. That’s a working attack.

Why Claude Mythos Won’t Be Released Publicly — Yet

Anthropic did not feel comfortable releasing this model to the general public. Logan Graham, who heads the team responsible for testing Anthropic’s AI model defenses, told CNN that “there’s a long way to go to have the appropriate safeguards.” This is actually the contradiction hiding inside the Glasswing story: Claude Mythos was not trained as a specialized security tool. It was built as a general-purpose model — one with exceptionally strong coding and reasoning skills for tasks like building AI agents and writing production code.

The fact that a general-purpose model turned out to be so dangerously capable in cybersecurity is precisely what changes everything. Sources suggest that OpenAI is finalizing a comparable model it will release only to a small set of companies through its existing “Trusted Access for Cyber” program. This means the arms race between frontier labs and defensive infrastructure is not a future scenario—it is happening right now, in controlled, invite-only settings across Silicon Valley.

The coding benchmark numbers are equally staggering and deserve more attention than they received. On the SWE-bench Verified benchmark—the industry’s toughest coding evaluation—Mythos solved 93.9% of all problems, compared to Claude Opus 4.6’s 80.8%. On the even harder SWE-bench Pro, Mythos hit 77.8% accuracy against Opus 4.6’s 53.4%. This isn’t just a security model. It’s a fundamentally more capable model across the board—which is precisely what makes it both extraordinary and risky. 

What Comes Next — and Why Claude Mythos Changes Everything

Industry insiders hint that the window for defenders to gain an advantage is extremely short. Alex Stamos, chief product officer at cybersecurity firm Corridor and former security chief at both Facebook and Yahoo, warned that “we only have something like six months before open-weight models catch up to the foundation models in bug finding—at which point every ransomware actor will be able to find and weaponize bugs without leaving traces for law enforcement.” That six-month window is the real story hiding beneath all the technical benchmarks.

Anthropic has already been briefing senior U.S. government officials—across the Cybersecurity and Infrastructure Security Agency, the Commerce Department, and the Center for AI Standards and Innovation—on Mythos’ full offensive and defensive cyber capabilities. The company has also made itself available for the government’s own testing and evaluation. According to reports, Anthropic privately warned top officials that Mythos makes large-scale AI-driven cyberattacks significantly more likely in 2026, which adds a geopolitical dimension to an already complex announcement.

When I first heard about this, I didn’t think much of it beyond standard frontier lab competition, but after digging into the specifics, I changed my mind completely. Anthropic’s goal is not to control Mythos forever. The company has stated clearly that its eventual aim is to “enable users to safely deploy Mythos-class models at scale — for cybersecurity purposes, but also for the wide range of other benefits that such highly capable models will bring.” The safeguards being developed today will eventually be ported to less powerful Opus models, allowing Anthropic to stress-test them at scale without exposing the full capability of Mythos to the open market.

The glasswing butterfly is transparent by nature. The vulnerabilities hiding inside global software have been invisible for years, surviving decades of human review, millions of automated tests, and entire generations of security infrastructure. That window of invisibility is closing fast. Right now, Claude Mythos is the most powerful tool available to close it—and the race to patch the world’s most critical systems before attackers get their hands on equivalent technology has already begun. The next twelve months will determine whether that head start was enough.