Anthropic confirmed the development of its most powerful model, Claude Mythos—details emerged from a leak.

Brief about the event

Anthropic officially confirmed that it is developing a new artificial intelligence model — Claude Mythos. It surpasses all previous versions in functionality and, as the company notes itself, carries “unprecedented” cybersecurity risks. Information about this emerged from a random leak of drafts that ended up in an open data cache.

1. What happened
DateEventBy Thursday eveningA draft blog post by Anthropic was found in public cache mentioning Claude Mythos as an AI creating “unprecedented” risks in the field of cybersecurity.After notifying FortuneThe company shut down public search on the cache and restricted access to documents.The cause of the leakAnthropic explains it was a “human misunderstanding” during content management system setup.

2. Scale of the leak
* About 3,000 items, including blog drafts and internal documents, were made publicly available.

* Among the leaked files:
* PDF about a closed event for CEOs of European companies in the UK (featuring Dario Amodei).
* Documents describing new levels of AI models — Capybara and Claude Mythos.

* The volume estimate was made by two experts:
* Alexander Povel – cybersecurity researcher from University of Cambridge.
* Roy Paz – senior AI security researcher at LayerX Security.

3. What are Capybara and Claude Mythos
ElementDescriptionCapybaraNew level of models, larger and more intelligent than Opus (the current “most powerful” in the lineup). Internally it belongs to the same base as Claude Mythos.Claude MythosActually the newest model from the company. It has significantly better results in programming, academic reasoning, and cybersecurity compared to Opus 4.6.
* Currently Anthropic sells three levels of models: Opus, Sonnet, and Haiku.

4. Key focus – cybersecurity risks
* The Claude Mythos model is described as a system that can identify vulnerabilities faster than traditional defense methods.

* According to the company’s plan, early access will be given only to organizations involved in cyber defense so they can prepare their systems for potential AI‑driven attacks.

5. Comparison with competitors
CompanyModelKey pointOpenAI GPT‑5.3‑Codex (Feb)First model classified as “Preparedness” in cybersecurity; trained to detect software vulnerabilities.Anthropic Opus 4.6 (today)Capable of discovering previously unknown vulnerabilities in codebases.

6. What’s next?
* Claude Mythos is at the final training stage and is planned for launch in limited early access.

* The company emphasizes that the model is not yet ready for mass use due to high operational costs.

Thus, the leak revealed details about Anthropic’s new model, its potential, and serious risks for the cybersecurity industry. The company has already taken steps to limit access to leaked materials and explained the cause of the incident.