Particle.news
Download on the App Store

Anthropic Confirms Leak of 'Claude Mythos,' Its Most Powerful AI Yet

Investor fears about advanced cyber capabilities sent major security stocks lower.

Overview

  • After a Friday report by tech outlet Decrypt on a leaked draft, Anthropic confirmed that “Claude Mythos” exists and called it the company’s most powerful model.
  • Leaked test results say Mythos beats Claude Opus 4.6 in coding tasks, academic reasoning, and cybersecurity evaluations.
  • The same files describe Mythos as a new tier above the current Opus line and note a follow-on model with the internal codename Capybara.
  • Draft language warns the model’s cyber skills could help attackers probe and exploit software faster than defenders can respond, so Anthropic plans an extremely cautious release.
  • U.S. cybersecurity stocks fell Friday on the news, with Palo Alto Networks down about 7%, CrowdStrike 6.4%, Zscaler 5.8%, and Fortinet 4%, echoing the sector’s selloff after February’s Claude Cowork launch.