Anthropic logo /Courtesy of Yonhap News

Artificial intelligence (AI) corporation Anthropic will first provide a preview version of its cutting-edge unreleased model "Mythos" to select corporations.

Anthropic on the 7th (local time) launched "Project Glasswing," aimed at strengthening cybersecurity, and said it will give priority access to some big tech corporations to "Claude Mythos Preview," the preview version of the top-tier model "Mythos" that has not yet been released to the public.

Mythos is a general-purpose AI model with strengths in identifying software vulnerabilities and security flaws, and will be provided on a limited basis only to corporations participating in Project Glasswing and to selected institutions. Amazon, Apple, Broadcom, Cisco, CrowdStrike, the Linux Foundation, Microsoft, and Palo Alto Networks are joining as Project Glasswing partners.

Anthropic said, "AI models have now reached a level that surpasses most people except top-tier experts in finding and exploiting vulnerabilities in software," and explained that it is initially restricting access to some corporations to prevent Mythos from being abused by malicious actors such as hacker groups.

Participating institutions will use "Claude Mythos Preview" to conduct security research, and Anthropic plans to strengthen the model's safety safeguards based on the results.

According to Anthropic, Mythos has detected thousands of zero-day vulnerabilities in recent weeks. Some were hard-to-find types, exemplified by a case where it discovered a 27-year-old bug in the OpenBSD operating system. It also found a vulnerability that had existed for 16 years in widely used video software, a flaw that had been missed even after running more than 5 million checks with an automated task tool.

Anthropic said, "We have no plans to release Claude Mythos Preview to the public," but added, "The ultimate goal is to safely deploy models at the Mythos level at scale."

Anthropic plans to offer participating corporations model usage rights worth up to $100 million and, separately, donate a total of $4 million to open-source security groups.

※ This article has been translated by AI. Share your feedback here.