- Claude Mythos leak reveals a new Capybara tier surpassing Opus models in reasoning, coding, and cybersecurity benchmarks.
- Anthropic warns that the model could enable advanced cyberattacks beyond current defensive capabilities if misused.
- Early access is limited to cybersecurity firms to prepare systems for emerging AI-driven exploit threats.
Claude Mythos’ major leak reveals Anthropic’s most advanced AI system yet. The model introduces a new capability tier while raising concerns over cybersecurity risks and controlled deployment.
Leak Exposes Claude Mythos and Capybara Tier
According to the First Post report, Claude Mythos entered public discussion after a data leak exposed internal draft materials. Nearly 3,000 unpublished assets were briefly accessible.
The exposed materials included blog drafts, PDFs, and internal references. Among them was detailed information about a new model tier called Capybara. This tier reportedly sits above Anthropic’s Opus models.
Capybara is described as a major upgrade in scale and performance. It reportedly delivers stronger results in reasoning, coding, and cybersecurity testing. Internal references suggest it surpasses Claude Opus 4.6 across several benchmarks.
Anthropic Confirms Testing and Controlled Access
Anthropic confirmed that Claude Mythos is undergoing early testing. A company spokesperson described it as the most capable system built so far and is limited to select early users.
The company clarified that core infrastructure and internal systems remained unaffected by human error in configuring the CMS.
These circulating excerpts aligned with the leaked draft content. They reinforced claims about the model’s advanced performance.
Cybersecurity Risks Drive Cautious Rollout
Claude Mythos has raised concerns due to its cybersecurity capabilities. Internal drafts describe the model as ahead of existing AI systems. It may identify and exploit software vulnerabilities at scale.
Anthropic acknowledged that bad actors could misuse these capabilities to launch large-scale AI-driven cyberattacks, posing a serious threat to current defense systems.
To address this, Anthropic created a rollout strategy and is prioritizing early access for cybersecurity organizations. These groups will use the model to strengthen systems against AI-driven exploits.
The company has previously dealt with malicious attempts involving its AI tools. Some incidents involved coordinated campaigns targeting multiple sectors. These included finance, technology, and government entities.
Enterprise Strategy and Industry Positioning
The leak also revealed plans for a private CEO event in Europe. The event will host business leaders for discussions on AI adoption. Demonstrations of unreleased Claude features are expected.
Dario Amodei is scheduled to attend the gathering. The event reflects Anthropic’s growing focus on enterprise clients. It also signals efforts to expand its influence in global markets.
Claude Mythos appears central to this strategy. The model’s capabilities may strengthen Anthropic’s position in enterprise AI. At the same time, controlled deployment suggests careful risk management.
The company continues to categorize its models into Opus, Sonnet, and Haiku tiers. However, Capybara introduces a new top layer. This shift indicates a broader evolution in Anthropic’s AI development roadmap.