Claude Mythos: the model Anthropic doesn't want to release

Anthropic confirmed: Mythos represents a step change in AI performance. It's the most capable model the company has ever built.

But it won't be released to the public. At least not anytime soon.

What is Mythos?

Internally, it's part of a new tier called "Capybara" — above Opus, being larger and more expensive than any previous model.

It's a general-purpose model with strong performance across all areas. But there's one detail that changed everything: exceptional capability in cybersecurity.

The capability that frightens

Mythos has found thousands of zero-day vulnerabilities in critical systems. Many of them critical. Without human guidance.

A concrete example: it discovered a 27-year-old vulnerability in OpenBSD — and allowed remote crash simply by connecting to the machine.

For perspective:

Opus 4.6: success rate near 0% in autonomous exploit development
Mythos Preview: running the same benchmark (Firefox 147 vulnerabilities), developed 181 functional exploits and gained register control in 29 additional attempts

These capabilities were not trained explicitly. They emerged as a consequence of general improvements in code, reasoning, and autonomy.

Why wasn't it released?

It's the first large LLM since GPT-2 in 2019 to have its general release delayed due to risk of social impact.

Anthropic believes that releasing the model openly before strengthening defenses on critical systems could be catastrophic.

Project Glasswing: The Alternative

Instead of releasing, Anthropic created Project Glasswing.

The model is available in private preview to approximately 40 companies, including AWS, Apple, Microsoft, Google, CrowdStrike, Cisco, Palo Alto Networks, JPMorgan Chase, NVIDIA and Linux Foundation.

Anthropic is donating $100 million in access credits for organizations to audit their systems.

Current Availability

Mythos is available as a closed preview on Amazon Bedrock and Google Cloud Vertex AI (Private Preview). There is no defined timeline for public availability.

Conclusion

Mythos is a model superior to Opus 4.6, with unprecedented offensive and defensive security capabilities. That's why it's under extremely restricted access.

It's not censorship. It's prevention.

Published on Hive.blog | #ArtificialInteligence