Anthropic’s most powerful unreleased AI model surfaced early after a data cache leak exposed internal material that was not meant to be public. Fortune reported that the Claude developer has trained a new model called Mythos, described internally as “by far the most powerful AI model we’ve ever developed.” The reference appeared in a draft blog post left inside an unsecured, publicly searchable cache, where cybersecurity researchers also found nearly 3,000 other unpublished assets.
A draft post exposed a new top-tier model line
After being contacted by Fortune, Anthropic confirmed the model exists and called it a “step change” in AI performance, adding that it is the company’s most capable system so far. Anthropic said the model is being tested with “early access customers” and attributed the leak to “human error” in its content management system.
The draft also introduced a new tier named Capybara, presented as larger and more capable than Anthropic’s current Opus family. According to the unpublished post, Capybara scores much higher than Claude Opus 4.6 on software coding, academic reasoning, and cybersecurity tests. That wording matters. It points to a model designed not just for general performance, but for domains that directly affect security-critical infrastructure.
Why the leak matters for crypto security
The strongest implication for the crypto sector comes from the draft’s cybersecurity language. It said the model “poses unprecedented cybersecurity risks,” a description with obvious relevance to blockchain security, smart contract auditing, and the constant race between attackers and defenders in DeFi.
The article tied that concern to events already unfolding this week. Ripple announced an AI-driven security overhaul for XRP Ledger after an AI-assisted red team found more than 10 vulnerabilities in its 13-year-old codebase. Ethereum launched a dedicated post-quantum security hub backed by eight years of research. Resolv’s stablecoin also lost its peg after an attacker exploited a minting contract that had no oracle checks and relied on single-key access control. Failures like that could be identified earlier by stronger AI systems, or exploited faster than defenders can react.
Decentralized AI projects face a higher benchmark
The leak also changes the conversation around AI-linked crypto assets. The report noted that Bittensor’s decentralized network recently released Covenant-72B, a model positioned against Meta’s Llama 2 70B. That launch helped trigger a 90% rally in TAO and pushed combined subnet token market capitalization to $1.47 billion.
If Anthropic’s claimed “step change” holds up, the performance gap between centralized AI labs and permissionless networks becomes harder to ignore. For decentralized AI projects, the challenge is no longer just building competitive models. It is doing so while the bar on capability, operating cost, and security expectations keeps rising. Anthropic said it is being deliberate about any release because the model is expensive to run and not ready for general availability. After Fortune reached out, the public data cache was removed.
The leak carries its own message as well. A company developing what it described as an AI model with unprecedented cybersecurity capability ended up exposing news of that model through a publicly searchable store because of human error.