Anthropic's Most Powerful Model Accidentally Leaked

In Brief

A significant leap in AI capability appears to be weeks away in new models from both Anthropic and OpenAI.

Details of an unreleased Anthropic model called Claude Mythos leaked from an an unsecured content management system. It is described as a tier above previous model, Opus, with unprecedented cybersecurity capabilities.

Meanwhile, OpenAI has finished pre-training its next model, codenamed Spud. 

 

What Happened

A Fortune exclusive revealed that Anthropic accidentally exposed details of an unreleased model through an unsecured content management system. Roughly 3,000 unpublished assets were accessible to anyone, including draft blog posts, internal images, and documents about the model and an invite-only CEO retreat Anthropic was planning in the United Kingdom.

The leaked drafts describe a model nicknamed "Mythos," sitting in a new tier above the Opus models. Anthropic confirmed the model is real, calling it "a step change" and "the most capable we've built to date," with "dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity" compared to Claude Opus 4.6.

The drafts also warn that Mythos is "currently far ahead of any other AI model in cyber capabilities" and "presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders." Anthropic blamed the easy access on "human error in the CMS configuration." At the same time, OpenAI has finished pre-training its next model, codenamed "Spud." OpenAI CEO Sam Altman told staff he expects a "very strong model" within weeks that can "really accelerate the economy."

SmarterX founder and CEO Paul Roetzer broke down what these new models could mean for business leaders on Episode 207 of The Artificial Intelligence Show.

The Key Numbers

• 3,000 - Approximate number of unpublished assets, including draft blog posts, internal documents, and details about Anthropic's unreleased "Mythos" model, that leaked.
• 6-9% - Immediate drop in major cybersecurity stocks after the leak revealed Mythos's highly advanced cyber capabilities.
• 6 to 12 - Months that frontier AI labs are typically operating ahead of the public's understanding of AI capabilities. 
• 24/7 - Around-the-clock frequency at which bad actors or competitors will soon be able to run AI agents to automatically discover and exploit vulnerabilities

Why the Leak Matters 

The real story is the capability jump. The leak revelas a model that represents a clear step change in what AI can do, with particular strength in cybersecurity, coding, and reasoning.

"The bigger models worry me," says Roetzer. "They talk specifically about reasoning, coding, and cybersecurity. None of this is new. We've known all the models are getting better at these things. But just the fact of how unprepared people are for what already exists, and to know we're very close to these next-level models, is worrying."

The timing tells you a lot. If Anthropic had draft blog posts, images, and marketing assets queued up in a CMS, the model is close to launch. These models likely finished training months ago and have been in post-training, red teaming, and safety evaluation since.

"There is always a more powerful model in training. The labs have always seen six to 12 months ahead of what you know to be true about reality."

Paul Roetzer, founder and CEO of SmarterX, Episode 207 of The Artificial Intelligence Show

Wall Street noticed immediately. The confirmation of a specific model with cyber capabilities was enough to move billions in market value in a single day. Cybersecurity stocks slumped on the news: CrowdStrike, Palo Alto Networks, and Zscaler each dropped about 6%. Okta and Netskope each fell more than 7%. Tenable plummeted 9%. The market reaction was swift, even though improved AI cyber capabilities have been predicted for years. 

Agents will make this kind of discovery routine. Roetzer pointed to a broader implication: as AI agents become more capable, competitive intelligence and vulnerability scanning will run 24/7. "You're trying to find vulnerabilities and exposures," he says. "You just run your agents 24/7 and go look for this kind of stuff." The era of accidental disclosure getting discovered by a reporter will give way to agents systematically finding every unsecured asset on the internet.

SmarterX Take

Anthropic's Mythos represents a leap in capability that the company itself describes as unprecedented in cybersecurity. OpenAI's Spud comes with Altman renaming the product organization to "AGI deployment." Both companies believe they are approaching something qualitatively different from what exists today.

For business leaders, the practical implication is straightforward: any AI strategy built around the current capabilities of these models has a short shelf life. Anthropic plans to release Mythos first to cybersecurity defense organizations before broader availability, which signals how seriously they take its potential. Organizations that have not started thinking about AI-powered threats to their systems should do so immediately.

What to Watch

Release of both models soon. Roetzer expects the launches are imminent, likely within weeks. The cybersecurity capabilities described in the Mythos drafts will reshape the threat landscape for every organization the moment the model becomes broadly available.

The broader pattern matters as much as the individual models. OpenAI's internal levels of AI progress are moving faster than anyone projected. Level one (chatbots) to the edge of level four (innovators) in roughly 20 months. Roetzer expects early signs of level five, where AI can do the work of an entire organization, to appear in some industries before the end of 2026.

Further Reading

Exclusive: Anthropic Left Details of an Unreleased Model in an Unsecured Data Trove → fortune.com

Anthropic Confirms 'Mythos' AI Model After Data Leak → fortune.com

Claude Mythos Details: Dramatically Higher Scores → the-decoder.com

Heard on The Artificial Intelligence Show, Episode 207
Paul Roetzer and Mike Kaput discuss what the Anthropic leak and pending release of OpenAI's Spud reveal about the next wave of AI capabilities. Listen →