Details of an unreleased Anthropic model called Claude Mythos became public in late March 2026 after a misconfigured content management system left thousands of the company’s internal files accessible. The exposed material included a draft announcement describing Mythos as Anthropic’s most capable model to date and warning that it could make cyberattacks more effective.
According to reports, security researchers found that roughly 3,000 files, including unpublished blog drafts and internal documents, could be reached without authentication. Anthropic removed public access after being alerted. The company has not disputed that the documents were genuine.
What the documents describe
The draft materials refer to a new product tier, reportedly codenamed Capybara, positioned above Claude Opus, which until then was the top of Anthropic’s lineup. The documents describe Mythos as performing substantially better than earlier Claude models on coding, reasoning and cybersecurity evaluations.
The most notable passages concern risk. In the leaked draft, Anthropic itself states that the model’s ability to find and exploit software vulnerabilities is advanced enough that it could outpace existing defensive tooling. The company writes that it is therefore planning a staged release, giving security organizations and defenders early access so they can harden their systems before the model becomes more widely available.
A security lapse at a safety-focused company
The exposure drew attention partly because Anthropic has built its public identity around AI safety and responsible deployment. Commentators noted the contrast between the company’s warnings about the model’s offensive cyber capabilities and the basic operational failure that let the warnings leak. Anthropic has not published a detailed account of how the misconfiguration occurred.
The incident follows earlier disclosures by Anthropic about attempts to misuse its models. The company has previously reported that groups used fraudulent accounts to extract outputs from Claude at scale, and in 2025 it described a case in which a state-linked actor used Claude Code to automate parts of an intrusion campaign. Those reports, together with the Mythos documents, reflect a view inside the company that frontier models have become a security concern in their own right, not only a technology.
What happens next
Anthropic has not announced a release date for Claude Mythos. Based on the leaked plan, the first external users are expected to be cybersecurity firms and other defenders rather than the general public. Markets reacted to the disclosure as well, with shares of some cybersecurity companies moving on speculation about how AI-driven vulnerability discovery could affect the industry.
The episode underscores a tension that is likely to recur as models grow more capable: the same systems that companies describe as too powerful to release carelessly are being developed inside organizations whose own security practices will be tested along the way.
