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

Claude creator Anthropic is creating a brand new AI mannequin known as Claude Mythos, described internally as the corporate’s most succesful mannequin so far, with draft supplies concerning the system being leaked on-line this week.

The existence of the mannequin was first reported by Fortune on Thursday after unpublished recordsdata tied to Anthropic’s weblog have been found in a publicly accessible information cache. An Anthropic spokesperson confirmed the existence of the mannequin to the publication.

“We’re creating a basic function mannequin with significant advances in reasoning, coding, and cybersecurity,” an Anthropic spokesperson instructed Fortune. “Given the energy of its capabilities, we’re being deliberate about how we launch it. As is normal observe throughout the trade, we’re working with a small group of early entry prospects to check the mannequin. We think about this mannequin a step change and probably the most succesful we’ve constructed so far.”

In an archived improvement page reviewed by Decrypt, Anthropic known as Mythos “probably the most highly effective AI mannequin we’ve ever developed.”

“Mythos is a brand new title for a brand new tier of mannequin: bigger and extra clever than our Opus fashions—which have been, till now, our strongest,” Anthropic wrote. “We selected the title to evoke the deep connective tissues that hyperlink collectively information and concepts.”

In accordance with Anthropic, Mythos scored “dramatically greater” than Claude Opus 4.6 on checks of software program coding, educational reasoning, and cybersecurity.

The leak of Mythos seems to have originated from draft supplies saved in an unsecured content material administration system. In accordance with Fortune, Anthropic restricted public entry to the information retailer after being notified that the recordsdata have been searchable on-line. The corporate attributed the publicity to human error within the configuration of its CMS instruments.

Nonetheless, Anthropic’s paperwork labeled Mythos as model one of many new mannequin, and described model two internally as “Capybara,” which the corporate additionally positioned above its present top-tier Opus fashions.

The draft supplies additionally highlighted issues concerning the system’s potential cybersecurity implications.

“Though Mythos is presently far forward of every other AI mannequin in cyber capabilities, it presages an upcoming wave of fashions that may exploit vulnerabilities in ways in which far outpace the efforts of defenders,” the corporate wrote.

Due to these dangers, the corporate stated it plans to launch the mannequin cautiously, starting with a restricted early-access rollout geared toward organizations engaged on cybersecurity protection.

Anthropic didn’t instantly reply to Decrypt’s request for remark.

Whereas Anthropic took down the weblog publish, information of the leak shortly spilled into monetary markets.

Shares of a number of cybersecurity corporations dropped after the stories surfaced, together with Palo Alto Networks (PANW), which fell about 7%, and CrowdStrike (CRWD), which dropped roughly 6.4%. In the meantime, Zscaler (ZS) declined round 5.8%, and Fortinet (FTNT) slipped about 4% throughout Friday buying and selling, in response to Yahoo Finance.

The selloff response echoes an analogous market response to the reveal of a brand new Anthropic product. In February, Anthropic unveiled Claude Cowork, an AI system designed to automate advanced office duties—together with contract evaluation and compliance—which triggered a broad sell-off throughout software program and professional-services firms.

That sell-off erased roughly $285 billion in market worth as traders reassessed the long-term influence of AI brokers on enterprise software program companies.

“The market’s response was a sign, not that AI brokers will instantly substitute these companies, however that traders are lastly pricing within the structural threat that basis mannequin suppliers can now compete instantly with the software program layer,” Nexatech Ventures founder Scott Dylan instructed Decrypt on the time. “That is a well mannered method of claiming if Anthropic can construct a authorized workflow device in-house, what’s stopping them from doing the identical for finance, procurement, or HR?”