Inside the Launch of Claude Mythos

Claude Mythos evolved from a leak suggesting Anthropic was testing a substantially more advanced model into a tightly restricted release defined by its exceptional cybersecurity capabilities.

Mythos surfaced not by launch but by leak: ~3,000 internal Anthropic files, including a draft blog, were briefly exposed in March 2026. Anthropic blamed human error and acknowledged the model.

Anthropic's most capable model yet — general-purpose, with a new architecture beyond the Claude 3 and 4 families and far stronger coding, reasoning, and security than Opus 4.6.

Mythos solves 73% of expert hacking challenges and finds zero-days. The UK AI Security Institute called it a major step forward in autonomous cyberattack capability.

AI-assisted attacks were already real — Chinese state actors, a Russian criminal hitting 600 devices via Claude and DeepSeek. Experts called agentic attackers a watershed event.

Anthropic gated Mythos through Project Glasswing — a $100M initiative — releasing it to ~40 organizations with no public date, deliberately limiting its vulnerability-finding power.

Washington split over access, the Fed convened bank CEOs, and cyber stocks fell. Yet Anthropic drew ~$800B offers as OpenAI and crypto exchanges raced to respond.


Opening sign-in…