On March 26, 2026, a misconfigured content management system at Anthropic accidentally exposed nearly 3,000 internal documents to the public — including draft blog posts announcing their next-generation AI model. The irony writes itself: the company that champions AI safety just had one of the biggest data leaks in AI history.
What Leaked: Claude Mythos (Capybara)
The exposed documents include two versions of the same draft blog post, differing only in the model’s name: "Mythos" in version 1 and "Capybara" in version 2. Both describe a new tier of AI model positioned above the current flagship Claude Opus 4.6.
In Anthropic’s own words from the draft: it is "by far the most powerful AI model we’ve ever developed" and represents a "step change" in capabilities.
The Capabilities
According to the leaked draft, Claude Mythos achieves dramatically higher scores than Opus 4.6 across multiple benchmarks:
- Software Coding: Significantly outperforms Opus 4.6 on code generation and debugging tasks
- Academic Reasoning: Higher scores across reasoning and problem-solving benchmarks
- Cybersecurity: Described as "far ahead of any other AI model in cyber capabilities" — this is the most alarming detail
The Cybersecurity Warning
The leaked draft explicitly warns that Mythos poses "significant" and "unprecedented" cybersecurity risks because of its advanced offensive capabilities. The model can rapidly find and exploit software vulnerabilities at a level no other AI system has demonstrated.
Anthropic’s planned response: a cautious rollout starting with cyber defenders (helping secure systems) before any broader release. The model is described as expensive to run and not yet ready for general availability.
The Double Failure
This incident is remarkable for two reasons:
| The Security Failure | Nearly 3,000 internal documents exposed via a misconfigured CMS — basic data governance that a company of Anthropic’s caliber should never get wrong |
| The Irony Failure | The company that warns the world about AI risk accidentally leaked details of a model it itself describes as potentially dangerous |
Market Reaction
The leak has sent ripples through the cybersecurity sector. CrowdStrike (CRWD) and other cybersecurity stocks saw movement as investors weighed the implications of AI models with offensive cyber capabilities entering the market.
For AI investors, the leak confirms what many suspected: the gap between current frontier models and the next generation is not incremental — it’s a step function. Companies with early access to Mythos-class models will have a significant competitive advantage.
Current Status
After Fortune broke the story, Anthropic removed public access to the data store and confirmed they are already testing Claude Mythos with a small group of early access customers. No public release date has been announced.
The Bottom Line
We are entering an era where AI models are becoming powerful enough that their very existence is a security concern. Anthropic’s leak didn’t just reveal a product roadmap — it revealed the growing tension between capability and safety that every frontier AI lab now faces.
As we push toward more agentic, more capable systems, the biggest risk may not be the AI itself. It may be the humans who build it.
Sources: Fortune (Exclusive), The Decoder, Futurism