Claude Mythos Leak Exposes Anthropic's Most Dangerous AI Model

Three thousand files left wide open on the internet. That's how the world discovered Claude Mythos — Anthropic's most powerful AI model to date. A data breach that felt like old-school hacking, except this time the victim was one of 2026's most technologically advanced companies.

🔓 The leak that changed everything

Anthropic never expected journalists and security researchers to discover their new model first. The company had left nearly 3,000 files — including a draft blog post about Claude Mythos — sitting in an unsecured data store with public access. The company blamed "human error" in configuring their content management system. The digital assets were set to public by default — a classic mistake that feels particularly ironic for a company worried about cyber threats.

Roy Paz from LayerX Security and Alexandre Pauwels from Cambridge University spotted and analyzed the leaked documents. After Fortune reached out, Anthropic immediately locked down the data store.

🚀 Claude Mythos: A new tier of AI capability

Claude Mythos isn't just another improved model — it launches an entirely new tier. Anthropic calls it "Capybara" and places it above their existing Opus, Sonnet, and Haiku models. The leaked materials describe Capybara as "larger and smarter than Opus models" with "dramatically higher scores on coding, academic reasoning, and cybersecurity tests" compared to Claude Opus 4.6.

What drives the performance leap

Anthropic describes Mythos as "the most capable AI model we have developed to date." A company spokesperson called it a "step change" — industry speak for something beyond incremental improvement. The model appears to have completed training and entered limited customer testing. Anthropic characterizes it as a "general purpose model with substantial advances in reasoning, coding, and cybersecurity."

⚠️ Unprecedented cyber capabilities

Here's where things get interesting — and concerning. The leaked blog post states that Mythos "presents unprecedented cybersecurity risks."

30 organizations already breached by Chinese groups using Claude Code

The company claims the model is "currently far ahead of any other AI model in cyber capabilities." Worse, it "foreshadows an incoming wave of models that can exploit vulnerabilities in ways that far exceed the efforts of defenders."

The careful rollout strategy

Precisely because of these risks, Anthropic plans an extremely cautious release. The initial plan provides early access only to organizations working on the defensive side of cybersecurity. "We want to act with extra caution and understand the risks it presents," the document states. The idea is to give cyber defenders an advantage over hackers.

"We're releasing it in early access to organizations, giving them a head start in improving their codebase resilience against the incoming wave of AI-driven exploits."
Anthropic draft blog post

📉 Market reaction tells the story

The Mythos revelation didn't stay in tech news. Markets reacted immediately — and negatively. Cybersecurity stocks like Palo Alto Networks, CrowdStrike, Zscaler, and Fortinet dropped up to 6%. The iShares Expanded Tech-Software Sector ETF (IGV) fell nearly 3%, while even Bitcoin got hit, sliding back to €66,000. Investors seem to be pricing in the risk that increasingly powerful AI systems could bring to the industry.

Previous incidents cast shadows

This isn't Anthropic's first rodeo with these issues. In November 2025, the company admitted that a Chinese state-sponsored group used Claude Code to breach about 30 organizations, including tech companies, financial institutions, and government agencies. Anthropic needed 10 days to detect and stop the operation — showing how difficult these systems are to control in practice.

🔍 What the leak reveals about Anthropic

Beyond Mythos, the breach exposed other intriguing details. Nearly 3,000 assets were publicly accessible, from images and banners to PDFs describing an invite-only CEO retreat in Britain.

Exclusive CEO Retreat

Documents describe a two-day retreat for European company CEOs at an 18th-century manor hotel in the English countryside, where Dario Amodei will attend.

There were even files describing employee "parental leave" — showing the CMS misconfiguration was extensive.

The irony runs deep

There's bitter irony in a company creating AI models with "unprecedented cyber risks" making such a basic security mistake. As Futurism noted, "let's hope the new model wasn't responsible for Anthropic's corporate blog security."

🤔 Questions for the future

The Claude Mythos case raises fundamental questions about the new generation of AI models. First, how realistic are the promises? OpenAI's GPT-5, released in August 2025, proved a major disappointment, falling short of expectations. Second, how can companies balance technological progress with safety? Anthropic seems to be trying to find the sweet spot, but the leak incident itself shows how difficult this is in practice.

The market reaction suggests investors are beginning to realize the real risks that advanced AI systems pose — not just as opportunities, but as threats to existing industries.

Claude Mythos might truly be the "step change" Anthropic promises. Or it could be another marketing exaggeration in the AI industry. What's certain is that the leak that revealed it put discussions on the table that the company obviously wasn't ready to have publicly — at least not yet.

Claude Mythos Anthropic AI leak Capybara AI cyber security AI model data breach artificial intelligence

Sources:

3000 Files Exposed: Inside Anthropic's Secret Claude Mythos AI Model Leak