Anthropic Confirms Testing New 'Mythos' AI Model After Accidental Data Leak

Anthropic left draft blog posts and roughly 3,000 unpublished assets exposed in a public data lake, according to Fortune, which reported the breach between March 26 and 31. Security researchers Roy Paz of LayerX Security and Alexandre Pauwels of the University of Cambridge independently discovered the exposed files. Among the documents was a draft describing a new AI model called "Mythos" — characterized internally as a "step change" in performance above Claude Opus 4.6, currently the company's most capable model. Anthropic acknowledged "human error" in its CMS configuration and removed public access after Fortune made contact.

What Do the Leaked Materials Reveal About Mythos?

The draft blog post described Mythos as a new tier above the existing Claude model lineup. Anthropic's current production hierarchy runs from Haiku 4.5 (the fastest, most affordable model) through Sonnet 4.6 to Opus 4.6 (the most capable). A model above Opus would represent Anthropic's most significant capability jump since launching Claude 3 in March 2024, which first established the three-tier structure that the company has maintained through subsequent generations.

No release date, pricing, or benchmark results appeared in the leaked materials — or at least none that Fortune reported. The description "step change" suggests a generational leap rather than an incremental update, though marketing language in unreleased drafts frequently overpromises relative to final performance. Anthropic has not commented beyond acknowledging the CMS error.

How Does This Fit the Current AI Model Race?

The leak arrives during the most compressed period of competition in AI model development. OpenAI released GPT-5.4 with a 1-million-token context window. Google shipped Gemini 3.1 with expanded multimodal capabilities. Meta continues to release Llama models under open weights. Each company is racing to establish its model as the default foundation for consumer-facing AI applications — from shopping assistants and travel planners to financial tools and code generation.

Anthropic has positioned itself as the safety-focused alternative, emphasizing Constitutional AI and responsible scaling policies. A major capability leap would test whether the company can maintain that positioning while matching competitors on raw performance. The accidental nature of the disclosure also raises questions about internal information security practices at a company that sells itself partly on trustworthiness — though CMS misconfigurations are among the most common causes of data exposure across the technology industry (Verizon Data Breach Investigations Report, 2025).

What Does More Powerful AI Mean for Consumers?

Each generation of AI models improves the accuracy and usefulness of consumer-facing tools. More capable models can process longer product descriptions, compare more options simultaneously, and handle multi-step tasks like coordinating travel bookings or analyzing financial products. GPT-5.4's 1-million-token context window, for example, allows a single conversation to hold the equivalent of several novels' worth of product information — enabling comparisons at a scale no human shopper could manage manually.