Anthropic is debuting a brand new AI mannequin as a part of a cybersecurity partnership with Nvidia, Google, Amazon Net Companies, Apple, Microsoft, and different corporations. Challenge Glasswing, because it’s referred to as, is billed as a method for big corporations, and probably even the federal government, to flag vulnerabilities of their techniques with just about no human intervention.
Anthropic is providing its launch companions entry to Claude Mythos Preview, a brand new general-purpose mannequin that it’s not at the moment planning to publicly launch on account of safety considerations. Newton Cheng, the cyber lead for Anthropic’s frontier pink staff, informed The Verge that the mannequin will ideally give cyber defenders a “head begin” towards adversaries. The companions will use the mannequin to investigate their system to identify high-stakes vulnerabilities and assist patch them up. Entry is restricted to maintain those self same adversaries from utilizing it to seek out weak factors and conduct assaults.
Although Claude Mythos Preview wasn’t particularly skilled for cybersecurity functions, Anthropic stated in a launch that the mannequin’s “robust agentic coding and reasoning expertise” are behind its cybersecurity advances. In an interview with The Verge, Newton Cheng, the cyber lead for Anthropic’s frontier pink staff, declined to share particular particulars of the mannequin’s cybersecurity successes past the corporate’s publicly-released examples, however Anthropic’s weblog publish stated that in latest weeks, Mythos Preview has flagged “1000’s of high-severity vulnerabilities, together with some in each main working system and net browser.” Anthropic’s weblog publish doesn’t point out retaining people within the loop for the mannequin’s cybersecurity sweeps; in reality, it highlights that the mannequin recognized vulnerabilities “and develop[ed] many associated exploits — solely autonomously, with none human steering.”
Claude Mythos Preview’s existence was first reported final month in a knowledge leak, which Anthropic attributes to human error. Dianne Penn, a head of product administration at Anthropic, informed The Verge in an interview that the corporate is “taking steps by way of solidifying our processes … That was not associated to software program vulnerabilities in any method.”
Mythos Preview can be privately out there to the corporate’s Glasswing companions, which additionally embrace JPMorgan Chase, Broadcom, Cisco, CrowdStrike, the Linux Basis, and Palo Alto Networks, plus about 40 different organizations that preserve or construct software program infrastructure. For now, Anthropic will assist subsidize the price of utilizing it. The corporate says it should commit as much as $100 million in utilization credit, plus $4 million in direct donations to the Linux Basis and the Apache Software program Basis, stated Cheng. In the long run, as Anthropic and different AI corporations face strain to show a revenue, this system may evolve right into a paid service that gives a brand new income stream — if it really works nicely sufficient for corporations to maintain utilizing it.
Regardless of its extremely public latest conflict with the Trump administration, Anthropic additionally stated within the launch that it has been in “ongoing discussions with US authorities officers about Claude Mythos Preview and its offensive and defensive cyber capabilities.” When The Verge requested what that meant, Penn confirmed that the corporate had “briefed senior officers within the US authorities about Mythos and what it could do,” and that the corporate remains to be “dedicated to working intently with all totally different ranges of presidency.” Cheng stated that although Anthropic is “engaged with” the federal government, he declined to talk to precisely who the corporate had briefed.
