Anthropic’s Mythos model, exposed by a public publishing cache — confirmed

An exposed publishing cache revealed a model announcement before Anthropic intended to make it. On 26 March 2026, Fortune reported unpublished material describing a new model called Mythos and a higher model tier discussed as Capybara. The report traced the discovery to security researchers who had found publicly accessible material associated with Anthropic’s publishing systems.

The important distinction is what escaped: draft descriptions and other unpublished assets, not the model’s weights. Fortune reported that Anthropic acknowledged a human error in its content-management configuration. The company removed public access to the searchable material after being contacted.

The leaked drafts contained strong claims about capabilities and cybersecurity risk. Those statements were evidence of Anthropic’s internal or planned public positioning, not an independent demonstration of every capability. The existence of a dramatic draft does not establish that its language has survived review, that its benchmarks are reproducible, or that the product will be released on the schedule readers infer.

A later development supplied an unusually clear outcome. TechCrunch reported Anthropic’s limited Mythos preview on 7 April as part of a defensive-security initiative. That established that the central model disclosure was real. The confirmed badge here covers that existence and subsequent preview, not the most expansive interpretations of what the model could do.

The episode was significant because the company’s own publishing workflow exposed a strategically important announcement while it was emphasizing careful handling of powerful AI. The public can understand that contradiction without being directed to the cache. Our links lead to Fortune’s reporting, The Verge’s coverage and the later preview report. No unpublished assets, private-event attendance details, model files or access instructions are reproduced. A verified Fortune X post documents the spread of the report, with its counts captured separately from the historical event date.