Anthropic, the company behind the generative artificial intelligence service Claude, has decided to provide its latest AI model exclusively to major big tech companies first.
Anthropic announced on Wednesday that it has formed "Project Glasswing," a joint cybersecurity initiative with major technology companies, and will provide early access to "Mythos," its most advanced model that has not yet been released to the general public. The project is named after the glasswing butterfly, whose transparent wings symbolize the goal of finding and fixing security vulnerabilities that are difficult to detect. Initial partners in this joint initiative include technology and security companies such as Amazon Web Services (AWS), Apple, Broadcom, Cisco, Google, Linux Foundation, Microsoft, Nvidia, and Palo Alto Networks, as well as financial firms including JPMorgan Chase. Access to Mythos will also be provided to more than 40 organizations responsible for critical software infrastructure.
Anthropic explained that Mythos is being released on a limited basis because its exceptional performance creates potential for misuse. On "CyberGym," a benchmark that measures the ability to reproduce cybersecurity vulnerabilities, Mythos scored 83.1 percent. This significantly exceeds the 66.6 percent scored by Anthropic's previous top model, "Opus 4.6." Mythos outperformed Opus 4.6 across most metrics, including various coding benchmarks, reasoning indicators, and information retrieval capabilities. Notably, on "Humanity's Last Exam" (HLE), a test comprising PhD-level expert questions across various fields, Mythos became the first AI model to break through the 50 percent barrier, scoring 56.8 percent without tool use.
Anthropic explained that it is providing the technology to major tech and security companies for defensive purposes first, as widespread distribution of such AI capabilities could be exploited by hackers and other malicious actors. Dario Amodei, CEO of Anthropic, wrote on social media platform X (formerly Twitter): "The risks of mishandling this are obvious, but if we get this right, there's a real opportunity to create an internet and a world that is fundamentally safer than before AI-based cyber capabilities emerged."