Anthropic debuts preview of powerful new AI model Mythos in new cybersecurity initiative

Anthropic connected Tuesday released a preview of its caller frontier model, Mythos, which it says volition beryllium utilized by a tiny coterie of spouse organizations for cybersecurity work. In a previously leaked memo, the AI startup called the exemplary 1 of its “most powerful” yet.

The model’s constricted debut is portion of a caller information initiative, dubbed Project Glasswing, successful which much than 40 spouse organizations volition deploy the exemplary for the purposes of “defensive information work” and to unafraid captious software, Anthropic said. While it was not specifically trained for cybersecurity work, the preview volition beryllium utilized to scan some first-party and open-source bundle systems for codification vulnerabilities, the institution said.

Anthropic claims that, implicit the past fewer weeks, Mythos identified “thousands of zero-day vulnerabilities, galore of them critical.” Many of the vulnerabilities are 1 to 2 decades old, the institution added.

Mythos is simply a general-purpose exemplary for Anthropic’s Claude AI systems that the institution claims has beardown agentic coding and reasoning skills. Anthropic’s frontier models are considered its astir sophisticated and high-performance models, designed for much analyzable tasks, including agent-building and coding.

The spouse organizations previewing Mythos see Amazon, Apple, Broadcom, Cisco, CrowdStrike, the Linux Foundation, Microsoft, and Palo Alto Networks. As portion of the initiative, these partners volition yet stock what they’ve learned from utilizing the exemplary truthful that the remainder of the tech manufacture tin payment from it. The preview is not going to beryllium made mostly available, Anthropic said.

Anthropic besides claims that it has engaged successful “ongoing discussions” with national officials astir the usage of Mythos, though 1 would person to ideate that those discussions are analyzable by the information that Anthropic and the Trump medication are presently locked in a ineligible battle aft the Pentagon labeled the AI laboratory a supply-chain hazard implicit Anthropic’s refusal to let autonomous targeting oregon surveillance of U.S. citizens.

News of Mythos was primitively leaked successful a information information incidental reported past period by Fortune. A draught blog astir the exemplary (then called “Capybara”) was near successful an unsecured cache of documents disposable connected a publically inspectable information lake. The leak, which Anthropic subsequently attributed to “human error,” was primitively spotted by information researchers. “‘Capybara’ 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,” the leaked papers said, adding aboriginal that it was “by acold the astir almighty AI exemplary we’ve ever developed,” according to the report. 

In the leak, Anthropic claimed that its caller exemplary acold exceeded show areas (like “software coding, world reasoning, and cybersecurity”) met by its presently nationalist models, and that it could perchance airs a cybersecurity menace if weaponized by atrocious actors to find bugs and exploit them (rather than hole them, which is however Mythos volition beryllium deployed).

Last month, the institution accidentally exposed astir 2,000 root codification files and implicit fractional a cardinal lines of codification via a mistake it made successful the motorboat of mentation 2.1.88 of its Claude Code bundle package. The institution past accidentally caused thousands of codification repositories connected Github to beryllium taken down arsenic it attempted to cleanable up the mess.