Anthropic didn’t plan to make headlines this way.
Details of what the company privately calls its most powerful model to date codenamed Mythos spilled onto the internet after a configuration mistake in the company’s content management system left internal documents exposed in a public repository. Security researchers spotted it first.
The leaked material included a draft blog post describing Mythos as a reasoning-heavy, compute-intensive language model with notably stronger coding and analytical skills than anything Anthropic has released before. Once researchers began archiving the content, Anthropic quietly shut down public access to the data store. The company later confirmed the model’s existence to Fortune, attributing the exposure to a CMS misconfiguration.
What makes Mythos different from Claude’s existing lineup isn’t just raw performance it’s the intended audience. Anthropic’s own draft language pointed directly at enterprise security teams, stating the company wants to “seed Mythos” among them first and has already been running tests with a select group of early-access customers.
The warning in that draft was very clear. Anthropic said it wants to be extra careful before sharing it with more people because of cybersecurity risks.
The reason is simple.
If models today can already find software problems and help people use them in the wrong way, then a much stronger model could make those risks even bigger.
One leaked detail drew particular attention a described capability for the model to autonomously identify and patch issues in its own code, sometimes referred to as “recursive self-fixing.” Security specialists noted this narrows the gap between human and machine software engineering in ways that worry both defenders and, inevitably, attackers.
Analysts are divided on what this means for the broader cybersecurity industry. Some see Mythos as a tool that could compress the time it takes to run threat hunts, triage incidents and probe for weaknesses. Others point out that earlier-generation models were quickly repurposed for malware development and Mythos sits several rungs above those.
Markets reacted before Anthropic said a word. Shares of major cybersecurity vendors including CrowdStrike, Palo Alto Networks and Fortinet dipped as investors tried to price in what a frontier model embedded inside security tooling might mean for competitive dynamics.
Anthropic hasn’t announced a release date. The draft itself noted that Mythos is expensive to run and the company is still working on efficiency before any general availability. Another version of the document reportedly names the model Capybara suggesting even the final branding isn’t settled yet.
