The Claude Mythos Leak: Why the Capybara Tier Was Withheld

In late March 2026, Anthropic didn’t announce their newest frontier model via a press release. It was discovered via a catastrophic error.

A misconfiguration in Anthropic’s content management system (CMS) publicly exposed nearly 3,000 internal documents. Within those files was the system card for a brand-new intelligence tier codenamed “Capybara.”

Sitting high above the Haiku, Sonnet, and Opus tiers, this new model was officially titled Claude Mythos. The fallout was immediate: within a single trading session, traditional cybersecurity providers lost $14.5 billion in market value as investors realized a single autonomous agent was capable of automating elite-level infrastructure security.

The ‘Glass Cannon’ Specs

Leaked documents and industry analysts point to an unprecedented architecture:

• 10-Trillion Parameters: It is arguably the largest sparse Mixture-of-Experts (MoE) configuration ever built.
• 1-Million Token Dense Context: Mythos can hold the equivalent of five full-length feature film scripts—or massive proprietary enterprise codebases—in memory with zero recall degradation.
• Adaptive Thinking Mode: Unlike legacy 2024 models, Mythos defaults to internal reasoning blocks. It iterates over multiple adversarial vulnerabilities internally before rendering an output, solving “wicked” level security logic previously untouchable by AI.

Project Glasswing

Claude Mythos was deemed too dangerous for a public API release. Instead of democratizing the model, Anthropic established Project Glasswing—named after the Greta oto butterfly known for its transparent wings. The metaphor is explicit: Mythos is designed to autonomously hunt for vulnerabilities that “hide in plain sight” inside critical global architecture.

Anthropic isolated the model behind a vetted consortium of 12 founding partners, including AWS, Apple, Broadcom, CrowdStrike, JPMorgan Chase, and Palo Alto Networks.

This is a defensive quarantine. The program grants these partners a strictly monitored “defensive head start,” utilizing Mythos to aggressively scan and patch their own internal systems and open-source networks before a rival model of equivalent intelligence falls into adversarial hands.

The Geopolitical Paradox

The handling of Mythos has generated a bizarre paradox within Washington.

On one side, the Treasury and the Fed view Project Glasswing as a vital early-warning system necessary to protect the U.S. economy from widespread cyber-collapse.

Simultaneously, the Pentagon recently labeled Anthropic a “supply chain risk” because the company refused to strip the ethical guardrails off its frontier models for use in autonomous military operations. Anthropic is effectively functioning as the most vital shield for U.S. infrastructure while simultaneously fighting the Department of Defense in the D.C. Circuit Court of Appeals.

The AI arms race has moved out of the chatbot phase and into national defense. Check out the embedded video above for the full teardown of the CMS leak.