MYTHOS

◆ CAPYBARA ◆

Claude · Fourth Tier · Leaked March 2026

Performance & Architecture

92.0% HumanEval
Python code generation

+20.2pp MATH Level 5
vs. Claude Opus baseline

59.4% GPQA Diamond
graduate-level reasoning

65.4%+ Terminal-Bench 2.0
"dramatically surpassed"

~10T Parameters (est.)
Mixture of Experts architecture

Tier IV Capybara class
above Haiku · Sonnet · Opus

What Happened

The Leak

A single CMS visibility toggle left public exposed ~3,000 internal Anthropic assets on March 26, 2026. Investigative reporter Bea Nolan of Fortune found the unsecured directory and authenticated the contents with two independent researchers.

"Human error in the CMS configuration." — Anthropic spokesperson

The Model

Claude Mythos (codename: Capybara) is a fourth, super-tier AI above Opus — designed for autonomous agentic ecosystems, competition-level math, and frontier software engineering. Anthropic engineers called it "by far the most powerful AI model we've ever developed."

Recursive development loop: built using Claude Opus 4.6, then folded back in to build its successors.

The Threat

Internal assessments warn Mythos is "currently far ahead of any other AI model in cyber capabilities." Anthropic's Frontier Red Team documented 500+ zero-day vulnerabilities found autonomously by an earlier Opus model. Mythos represents a qualitative leap beyond that baseline.

"Presages an upcoming wave that can exploit vulnerabilities far outpacing defenders."

Timeline

Mar 26

CMS Exposure Discovered

Bea Nolan finds the open directory. Two researchers independently verify. Anthropic notified and locks it down within hours.

Mar 26–27

Fortune Report Published

The exclusive publishes overnight. Mythos benchmark data and threat warnings enter public circulation. Pentagon injunction won simultaneously.

Mar 27

$14.5B Market Crash

Cybersecurity sector flash crash. PANW −7.5%, CRWD −7.5%, TENB −11%. Global X IHAK ETF hits multi-year lows.

Mar 28

Anthropic Confirms

Mythos confirmed as real — "step change, most capable we've built." IPO discussions with Goldman, JPMorgan, Morgan Stanley reported.

"While Mythos currently far surpasses any other AI model in cybersecurity capability, it foreshadows an incoming wave where models will be able to exploit vulnerabilities at a rate far outpacing defenders' efforts."
— Internal Anthropic draft, recovered from CMS data lake, March 2026

🧠 Full Report ↗ ← Index