Due to the concerns of cybersecurity, Anthropic is controlling the access of a small number of partners while testing a new powerful AI model called Claude Mythos Preview.
A configuration error in the company’s internal content system in late March initially exposed the model, which is referred to as a general-purpose frontier system. In accordance with the information that is not yet made public, Claude Mythos will be a new catagory of devices with significantly larger number of features than Opus, which is the current flagship of the business.
Instead of making the model available to the general public, Anthropic is granting access through a project known as Project Glasswing. Just a handful of companies are going to utilize this model for defensive security tasks under this program.
Amazon, Apple, Broadcom, Cisco, CrowdStrike, Linux Foundation, Microsoft, Palo Alto Networks, as well as forty more businesses are going to be launch partners.
These participants will use open-source tools and the methodology to scan and safeguard their systems.
Claude Mythos performs well on coding and reasoning problems even though it was not created with cybersecurity in mind. The model earned 83.1% on the CyberGym vulnerability analysis benchmark, whereas Opus 4.6, which had previously topped the rankings, scored 66.6%.
According to Anthropic, the model has already found thousands of zero-day vulnerabilities, including serious problems. A previously unidentified problem in OpenBSD that has gone unpatched for 27 years was one example given. Additionally, the model showed how to combine many Linux kernel flaws to obtain superuser access.
Anthropic stated that it is releasing the model cautiously due to possible short-term cybersecurity threats. Before comparable capabilities become generally accessible, the business wants to give defenders time to fortify systems.
AI has greatly reduced the time between finding and exploiting vulnerabilities, according to CrowdStrike CTO Elia Zaitsev, who also noted that these techniques can function at scale.
Open-source ecosystems, where security resources are frequently scarce, are another target of the program. Jim Zemlin claims that by using AI, maintainers can easily find and address issues.
Anthropic is contributing $4 million to open-source security organizations along with $100 million in use credits to enterprises that are taking part.
Anthropic gave a statement in which they said that it intends to make Mythos-class models more widely available in the future. As the business assesses risks and possible effects, access is now restricted to Project Glasswing participants.
