Anthropic is developing a new artificial intelligence model called Claude Mythos, according to a draft blog post that was inadvertently made public. The document describes the model as by far the most powerful system the company has built to date. The leak has drawn attention both to the model’s capabilities and to the circumstances under which the information became public.
Claude Mythos is said to belong to a new tier internally referred to as Capybara. According to the leaked draft, the model delivers dramatically improved performance compared to Claude Opus 4.6 across several key domains. These include coding, academic reasoning, and cybersecurity-related tasks.
The reported advances in cybersecurity performance have prompted concern among observers about the potential for misuse. Analysts note that a model with significantly enhanced capabilities in this area could introduce unprecedented risks if deployed without adequate safeguards. The dual-use nature of such technology means it can serve both defensive and offensive purposes.
Particular concern has been raised around areas such as DeFi security, where advanced AI tools could be exploited to identify and take advantage of vulnerabilities in decentralized financial systems. The combination of improved reasoning and cybersecurity proficiency makes the model a potentially powerful instrument in the wrong hands. These risks are seen as compounding the broader challenges already associated with frontier AI development.
Adding a layer of irony to the situation, the leak itself resulted from what is being described as a basic content management error on Anthropic’s part. The company, which publicly emphasizes its commitment to cutting-edge cybersecurity capabilities, inadvertently exposed sensitive internal details through an operational oversight. The incident has prompted questions about internal information handling practices at one of the AI industry’s leading safety-focused organizations.
Originally reported by CoinDesk.
