The first public look at Claude Mythos, an unreleased AI model, came through a leak. Anthropic, the company behind the Claude AI models, had left drafts open in the software it uses to publish web pages. Fortune reported that close to 3,000 unpublished files linked to the company’s site were left in public view. The company said human error made them visible.
Those drafts gave the story its force. They described Mythos as Anthropic’s most powerful model yet, a “step change” in capability with meaningful advances in reasoning, coding and cybersecurity.
Anthropic’s public record said less. In March 2026, the company still presented Claude Opus 4.6 as its top released model. Its system cards, the public documents it publishes about its models, did not include Mythos. There was no public page or test result that would let outsiders judge how much stronger it was.
A cyber warning pushed the story beyond normal product chatter. The draft said the model could pose serious cybersecurity risks, including the ability to find security flaws in software and exploit them faster than defenders could respond.
How polished the drafts looked split opinion. The material read like planned launch content, with structured web copy, headings and a publication date. Some saw it as almost too neat, like a press release that escaped early. Still, polished does not mean staged. External researchers found the material, and the company acknowledged the leak and called the files early drafts. A staged leak is possible in theory, but it is hard to square with a company deliberately leaking its own warning about a dangerous model.
Online, people ran ahead of those facts. They turned leaked language into broad claims about immediate cyber disruption, without the detail to support them. The leak suggested the company was working on something serious. It also suggested the company wanted the market to see that work as a major leap. Those are two different things.
Mythos may be materially stronger than Opus 4.6. At the time of the leak, though, its two biggest claims, the step change and the cyber risk, rested only on the company’s own drafts. No public tests, documentation or formal release stood behind them.
