Claude Mythos: The Most Powerful AI Model in the World That Anthropic Won't Release

What if an AI could find security flaws that the world's best hackers never detected? That's exactly what Claude Mythos Preview does — the most powerful model ever created by [Anthropic](https://anthropic.com). Announced on April 8, 2026, Mythos discovered thousands of zero-day vulnerabilities — some dating back 27 years — in every major operating system and web browser. Faced with this unprecedented power, Anthropic made a radical decision: not to release it publicly. Instead, the company launched Project Glasswing, an elite consortium bringing together AWS, Apple, Google, Microsoft, Nvidia, and other tech giants to use Mythos defensively.

What is Claude Mythos?

Claude Mythos — from the Ancient Greek meaning "utterance" or "narrative" — is a general-purpose frontier language model. It's not specialized in cybersecurity: it excels across all benchmarks. But it's in discovering and exploiting software vulnerabilities that it stands out spectacularly. According to Anthropic, Mythos Preview can "surpass all but the most skilled humans" at detecting security flaws.

The model is the result of Anthropic's most ambitious research program. With an internal codename of "Capybara", Mythos first leaked to the press in late March via Fortune, before Anthropic officially confirmed its existence on April 8 with the launch of Project Glasswing.

Unprecedented cybersecurity capabilities

In just a few weeks of testing, Mythos Preview identified thousands of zero-day vulnerabilities — flaws previously unknown to the developers of the affected software. Many are critical. Some had survived decades of human review and millions of automated tests. Three striking examples:

• OpenBSD — 27-year-old vulnerability: Mythos found a flaw in OpenBSD, one of the most security-hardened operating systems in the world, used for firewalls and critical infrastructure. The vulnerability allowed an attacker to remotely crash any machine simply by connecting to it.
• FFmpeg — 16-year-old vulnerability: In a line of code that automated testing tools had hit five million times without ever catching the problem. FFmpeg is used by countless software to encode and decode video.
• Linux kernel — exploit chain: Mythos discovered and chained together several vulnerabilities in the Linux kernel — the software that runs most of the world's servers — to allow full privilege escalation from ordinary user access to complete machine control.

Benchmarks: Mythos crushes the competition

Mythos Preview doesn't just dominate in cybersecurity. It's the top-performing model on virtually every frontier benchmark:

The most striking gap is on SWE-bench Pro — the real-world software problem-solving benchmark — where Mythos reaches 77.8% compared to 53.4% for Opus 4.6. That's a nearly 25-point lead, a gap rarely seen between two model generations.

Project Glasswing: a consortium to secure the world

Faced with Mythos's power, Anthropic chose caution. Rather than making the model available to everyone, the company created Project Glasswing — named after the glasswing butterfly Greta oto — an unprecedented defensive consortium in AI history.

Founding partners include the biggest names in tech and cybersecurity:

• Amazon Web Services — Cloud stack defense
• Apple — iOS/macOS ecosystem security
• Google — Security tools like Big Sleep and CodeMender
• Microsoft — Cybersecurity research and CTI-REALM benchmark
• Nvidia — GPU/AI infrastructure security
• Cisco — Critical network protection
• CrowdStrike — Threat detection and response
• JPMorgan Chase — Financial system resilience
• Palo Alto Networks — Cybersecurity stack modernization
• Linux Foundation — Open source security

In total, over 50 organizations have access to Mythos Preview. Anthropic has committed $100 million in usage credits and $4 million in donations to open-source security organizations (Linux Foundation, Apache Software Foundation).

“AI capabilities have crossed a threshold that fundamentally changes the urgency required to protect critical infrastructure from cyber threats, and there is no going back. The old ways of hardening systems are no longer sufficient.”

Why Mythos isn't available to the public

Anthropic's decision is driven by a simple observation: the same capabilities that make Mythos an extraordinary defensive tool also make it a potentially devastating weapon in the wrong hands. The cost, effort, and expertise required to find and exploit vulnerabilities have dropped dramatically with the latest frontier models.

Anthropic plans to develop cyber safeguards that can detect and block dangerous uses, test them first on less powerful models (like Opus 4.7), and then gradually open Mythos-class models to the public. But this process will take time.

What Mythos means for the future of AI

Claude Mythos marks a turning point. For the first time, an AI model surpasses the best humans in both offensive and defensive cybersecurity. The Council on Foreign Relations calls it an "inflection point for AI and global security". Bruce Schneier, a world-renowned cryptography expert, considers Project Glasswing a necessary response.

For Claude users, the good news is that Opus 4.7, while less powerful than Mythos, already integrates security lessons learned from Mythos. And future generations of Claude models will benefit from safeguards tested and refined through Glasswing.

Frequently asked questions