Anthropic CEO and co-founder Dario Amodei speaks throughout the 56th annual World Financial Discussion board assembly in Davos, Switzerland, Jan. 20, 2026.
Denis Balibouse | Reuters
Anthropic on Tuesday introduced a sophisticated synthetic intelligence mannequin that may roll out to a choose group of firms as a part of a brand new cybersecurity initiative referred to as Undertaking Glasswing.
The mannequin, Claude Mythos Preview, excels at figuring out weaknesses and safety flaws inside software program, and Anthropic is limiting entry to attempt to stop dangerous actors from exploiting that functionality, the corporate stated.
Anthropic stated Apple, Google, Microsoft, Nvidia and Amazon Internet Providers are among the many challenge’s preliminary launch companions and can be capable of use the mannequin for defensive safety work. Greater than 40 different firms, together with CrowdStrike and Palo Alto Networks, are additionally taking part, Anthropic stated.
“There was lots of inside deliberation,” Dianne Penn, Anthropic’s head of analysis product administration, instructed CNBC in an interview. “We actually do view this as a primary step for giving lots of cyber defenders a head begin on a subject that shall be more and more essential.”
Anthropic’s announcement comes after descriptions of the mannequin have been found by Fortune in a publicly accessible knowledge cache late final month. Cybersecurity shares fell on the report, which stated that the mannequin had superior cyber capabilities that additionally posed a big danger.
The iShares Cybersecurity ETF was largely flat throughout intraday buying and selling on Tuesday.
“The risks of getting this flawed are apparent, but when we get it proper, there’s a actual alternative to create a basically safer web and world than we had earlier than the arrival of AI-powered cyber capabilities,” CEO Dario Amodei wrote in a submit on X touting the rollout of Undertaking Glasswing.
Anthropic was based in 2021 by a bunch of researchers and executives who defected from OpenAI over issues about its path and perspective towards security.
The corporate spent years rigorously developing its fame as a agency that was extra devoted to accountable AI deployment, and it unveiled Undertaking Glasswing simply weeks after its high-profile conflict over security with the Protection Division spilled into public view.
Anthropic stated it has been in “ongoing discussions” with U.S. authorities officers about Claude Mythos Preview’s cyber capabilities.
Staff at Anthropic selected the title Undertaking Glasswing, a metaphor that likens a clear butterfly to software program vulnerabilities, that are “comparatively invisible,” Penn stated.
Claude Mythos Preview is ready to discover bugs, a few of that are vital, which have beforehand been troublesome to detect, Anthropic stated. In a single case, the corporate stated, the mannequin recognized a 27-year outdated bug in OpenBSD, which, in line with its web site, is an working system that emphasizes safety.
Anthropic stated Claude Mythos Preview is a general-purpose mannequin that was not particularly educated for cybersecurity, and its improved cyber capabilities are a results of its robust coding and reasoning abilities. The corporate stated it doesn’t plan to make the mannequin typically accessible, however the purpose is to study the way it might finally deploy Mythos-class fashions at scale.
Firms that take part in Undertaking Glasswing all construct or keep vital software program infrastructure, and Anthropic stated they are going to use its fashions to attempt to safe each first-party and open-source techniques. Anthropic stated it has dedicated as much as $100 million in utilization credit for these efforts, however companions pays to make use of the mannequin previous that threshold.
Newton Cheng, Anthropic’s Frontier Crimson Staff cyber lead, stated the corporate desires the corporations to get used to leveraging these capabilities earlier than they change into broadly accessible. The corporate stated it is attempting to keep away from “recklessly or irresponsibly” releasing a mannequin that adversaries might reap the benefits of.
“Cybersecurity is simply going to be an space the place this broad improve in capabilities has potential for danger, and thus we now have to maintain a extremely shut eye on what is going on on there,” Cheng stated in an interview.
WATCH: Panic on cyber shares feels overdone, says Massive Expertise’s Alex Kantrowitz