Anthropic's 'Most Capable' AI Model Claude Mythos Leaks, Deemed Major Cybersecurity Threat

Claude creator Anthropic is processing a caller AI exemplary called Claude Mythos, described internally arsenic the company’s astir susceptible exemplary to date, with draught materials astir the strategy being leaked online this week.

The beingness of the exemplary was archetypal reported by Fortune connected Thursday aft unpublished files tied to Anthropic’s blog were discovered successful a publically accessible information cache. An Anthropic spokesperson confirmed the beingness of the exemplary to the publication.

“We’re processing a wide intent exemplary with meaningful advances successful reasoning, coding, and cybersecurity,” an Anthropic spokesperson told Fortune. “Given the spot of its capabilities, we’re being deliberate astir however we merchandise it. As is modular signifier crossed the industry, we’re moving with a tiny radical of aboriginal entree customers to trial the model. We see this exemplary a measurement alteration and the astir susceptible we’ve built to date.”

In an archived improvement page reviewed by Decrypt, Anthropic called Mythos “the astir almighty AI exemplary we’ve ever developed.”

“Mythos is simply a caller sanction for a caller tier of model: larger and much intelligent than our Opus models—which were, until now, our astir powerful,” Anthropic wrote. “We chose the sanction to evoke the heavy connective tissues that nexus unneurotic cognition and ideas.”

According to Anthropic, Mythos scored “dramatically higher” than Claude Opus 4.6 connected tests of bundle coding, world reasoning, and cybersecurity.

The leak of Mythos appears to person originated from draught materials stored successful an unsecured contented absorption system. According to Fortune, Anthropic restricted nationalist entree to the information store aft being notified that the files were searchable online. The institution attributed the vulnerability to quality mistake successful the configuration of its CMS tools.

However, Anthropic’s documents labeled Mythos arsenic mentation 1 of the caller model, and described mentation 2 internally arsenic “Capybara,” which the institution besides positioned supra its existent top-tier Opus models.

The draught materials besides highlighted concerns astir the system’s imaginable cybersecurity implications.

“Although Mythos is presently acold up of immoderate different AI exemplary successful cyber capabilities, it presages an upcoming question of models that tin exploit vulnerabilities successful ways that acold outpace the efforts of defenders,” the institution wrote.

Because of those risks, the institution said it plans to merchandise the exemplary cautiously, opening with a constricted early-access rollout aimed astatine organizations moving connected cybersecurity defense.

Anthropic did not instantly respond to Decrypt’s petition for comment.

While Anthropic took down the blog post, quality of the leak rapidly spilled into fiscal markets.

Shares of respective cybersecurity firms dropped aft the reports surfaced, including Palo Alto Networks (PANW), which fell astir 7%, and CrowdStrike (CRWD), which dropped astir 6.4%. Meanwhile, Zscaler (ZS) declined astir 5.8%, and Fortinet (FTNT) slipped astir 4% during Friday trading, according to Yahoo Finance.

The selloff absorption echoes a akin marketplace effect to the uncover of a caller Anthropic product. In February, Anthropic unveiled Claude Cowork, an AI strategy designed to automate analyzable workplace tasks—including declaration reappraisal and compliance—which triggered a wide sell-off crossed bundle and professional-services companies.

That sell-off erased astir $285 cardinal successful marketplace worth arsenic investors reassessed the semipermanent interaction of AI agents connected endeavor bundle businesses.

"The market's effect was a signal, not that AI agents volition instantly regenerate these businesses, but that investors are yet pricing successful the structural hazard that instauration exemplary providers tin present vie straight with the bundle layer," Nexatech Ventures laminitis Scott Dylan told Decrypt astatine the time. “That's a polite mode of saying if Anthropic tin physique a ineligible workflow instrumentality in-house, what's stopping them from doing the aforesaid for finance, procurement, oregon HR?”