Anthropic, the AI safety-focused company behind the Claude family of models, has confirmed that it is developing and testing a powerful new AI model after an accidental data leak revealed its existence. The leak occurred when draft content was left in a publicly accessible data store due to a configuration error in the company's content management system. Cybersecurity researchers discovered the material, which included a draft blog post describing the new model in detail. Anthropic acknowledged the incident, calling it a result of human error, and quickly removed public access to the exposed data.
Meet Claude Mythos and the Capybara Tier
The leaked draft blog post revealed that the new model is called Claude Mythos. Anthropic described it internally as the most powerful AI model the company has ever developed. The document also introduced a new model tier called Capybara, which sits above the existing Opus tier in terms of size and capability. Currently, Anthropic offers its models in three tiers — Opus being the largest and most capable, followed by Sonnet, and then Haiku as the smallest and fastest. Capybara would represent a fourth and even more powerful tier, though it would also be more expensive to run.
A Step Change in AI Performance
Anthropic confirmed the model's existence in an official statement, describing it as a general-purpose model with meaningful advances in reasoning, coding, and cybersecurity. The company called it a step change in capability and the most capable model it has built to date. According to the leaked document, the new model scores dramatically higher than Claude Opus 4.6 on benchmarks related to software coding, academic reasoning, and cybersecurity tasks. The model has already been made available to a small group of early access customers for testing.
Serious Cybersecurity Concerns
One of the most striking revelations from the leak involves the model's cybersecurity implications. According to the draft blog post, the new model is currently far ahead of any other AI model in cyber capabilities. Anthropic expressed concern that hackers could potentially use the model to run large-scale cyberattacks and exploit software vulnerabilities faster than defenders can fix them. The company warned that this model signals an upcoming wave of AI systems capable of outpacing human cybersecurity efforts. Because of these risks, Anthropic plans to release the model first to cybersecurity organizations, giving defenders a head start in strengthening their systems before the model becomes widely available.
An Industry-Wide Challenge
Anthropic is not the only company facing these concerns. Earlier this year, a rival AI lab released a model that it classified as high capability for cybersecurity-related tasks under its own safety framework. That model was directly trained to identify software vulnerabilities. Anthropic's own Opus 4.6, released around the same time, also demonstrated the ability to discover previously unknown vulnerabilities in production code. These developments highlight a growing industry-wide challenge where the same AI capabilities that help defend systems can also be weaponized by malicious actors.
The Leak Also Exposed Private Company Plans
The data leak went beyond the model announcement. Nearly 3,000 unpublished assets linked to Anthropic's blog were found to be publicly accessible. Among these were internal documents, including details about an exclusive invite-only CEO summit planned in the United Kingdom. The event is described as an intimate two-day retreat at an 18th-century English countryside manor, where European business leaders would meet with Anthropic's CEO Dario Amodei to discuss the future of AI and experience unreleased Claude capabilities. The leak also reportedly exposed employee-related documents, raising additional privacy concerns.
A Cautious Rollout Strategy
Despite the leak forcing an early disclosure, Anthropic appears committed to a careful and measured release strategy for the new model. The draft blog post outlined plans to begin with a limited group of early access users before any broader rollout. The company emphasized that the model is expensive to run and not yet ready for general availability. This cautious approach reflects Anthropic's broader philosophy of prioritizing safety alongside capability, especially as AI models become increasingly powerful and potentially dangerous in the wrong hands.
