Back to Home

Claude Mythos: Leak and Cyber Power of Anthropic

Anthropic develops Claude Mythos — Capybara-level model surpassing Opus in key tests. CMS draft leak revealed details. Focus on cybersecurity with limited release for defenders.

Claude Mythos Leak: New Era of AI in Cybersecurity
Advertisement 728x90

# Anthropic Data Leak: Claude Mythos Changes the Game in AI for Cybersecurity

Anthropic is testing Claude Mythos—a new AI model positioned as the most powerful in the Claude lineup. The information leaked through a public draft of a blog post in the company's CMS. The model introduces the Capybara level, surpassing Opus in performance on programming, academic tasks, and cybersecurity.

Leak Details and Model Configuration

The leak occurred due to incorrect CMS settings: drafts were becoming public by default without manual privacy configuration. As a result, access was granted to 3000 files—posts, images, PDFs, including materials from a summit for European CEOs with Dario Amodei. Anthropic classified the incident as a human error, unrelated to AI systems.

Claude Mythos demonstrates significant progress over Claude Opus 4.6. The draft highlights superiority in coding tests, reasoning, and vulnerability detection. A company representative confirmed the development, noting a "qualitative leap".

Google AdInline article slot

Cybersecurity Capabilities and Risks

Key focus is on its cyber potential. Anthropic predicts that Mythos outperforms existing AIs in exploiting vulnerabilities, potentially outpacing defensive measures. The launch will begin with limited access for cybersecurity organizations to give defenders an edge.

The industry is noting rising threats:

  • In February, OpenAI rated GPT-5.3-Codex as High in cybersecurity.
  • Claude Opus 4.6 detects unknown vulnerabilities in code.
  • In November 2025, a Chinese group used Claude Code for autonomous attacks on 30 organizations (80–90% of operations without intervention).

Claude Levels Comparison and Trends

| Level | Key Strengths | Status |

Google AdInline article slot

|--------------------|--------------------------------|------------|

| Opus 4.6 | Coding, vulnerabilities | Released |

| Capybara (Mythos) | Programming, reasoning, cyber | Testing |

Google AdInline article slot

Trend: AI models are shifting from passive analysis to active exploitation. Anthropic warns of upcoming systems that accelerate both attacks and defenses.

Developers should consider these shifts when integrating AI into security pipelines. The limited release of Mythos will allow testing the model in a controlled environment.

Key Takeaways

  • Claude Mythos (Capybara) is Anthropic's first model surpassing Opus 4.6 in cyber tests.
  • CMS leak exposed 3000 private files, including the roadmap.
  • Priority access for cybersecurity pros minimizes risks.
  • Noted rise in AI attacks: 80–90% autonomy in real incidents.
  • Industry prepares for an era of AI outpacing human patching speeds.

— Editorial Team

Advertisement 728x90

Read Next