Anthropic leaked the strongest AI model from an unsecured cache.

Anthropic is testing its most powerful AI model to date, but the world only learned about it after a data leak. According to reports, a draft article from the company was left in an unprotected public data cache, revealing the name of the new model and the accompanying technical ambitions.

The incident began when a draft article was found in a publicly searchable data repository, alongside thousands of other unpublished assets. After Fortune reached out, Anthropic confirmed the existence of this model, stating that it represents a “breakthrough” in AI capability and is currently being tested with early access customers. The company also acknowledged a human error in its content management system that led to the incident.

The draft reveals a new model layer named Capybara, described as larger and more powerful than the previous Opus models — which were once Anthropic’s strongest versions. Compared to Claude Opus 4.6, Capybara is said to deliver superior results in software programming tests, academic reasoning, and cybersecurity.

The most noteworthy aspect for the crypto industry is the cybersecurity angle. The draft states that this model could create “unprecedented cybersecurity risks,” and that is a direct signal for the blockchain security field, smart contract auditing, and the ongoing arms race between attackers and defenders in DeFi. As stronger AI continues to emerge, both attackers and defenders will have better, faster, and more dangerous tools.

This week, Ripple also announced an AI-based security overhaul for the XRP Ledger after an AI-supported red team discovered over 10 vulnerabilities in the 13-year-old codebase. Ethereum recently opened a dedicated post-quantum security center. Conversely, the Resolv stablecoin lost its peg after an attacker exploited a mint contract lacking oracle checks and featuring only single-key access control — precisely the type of flaw that more advanced AI tools could detect before being exploited.

For the AI token market, the leak raises another question: as a concentrated lab like Anthropic makes such a capability leap, how much will the gap widen between well-funded enterprise models and decentralized AI projects? Bittensor recently launched Covenant-72B, causing TAO to surge, but the “breakthrough” from a top lab like Anthropic will clearly raise the competition standard to a new level.

Anthropic stated that it is being cautious with its model release plans due to high operational costs and that it is not yet ready for mainstream users. The company also removed public access to the data cache after being contacted by Fortune. Thus, this incident is both a story about an ambitious new AI model and a wry reminder: even the company developing the most advanced cybersecurity tools can make very basic security mistakes.