Anthropic Launches Cyber Alliance with AI Model 'Mythos'

Anthropic announced on the 7th (local time) its plan to launch 'Project Glasswing,' a cybersecurity collaboration with major technology and security firms. The name 'Project Glasswing' is inspired by the Glasswing butterfly, known for its transparent wings, signifying the initiative's aim to find and resolve hard-to-detect security vulnerabilities.

Initial partners include major tech and security companies such as Amazon Web Services (AWS), Apple, Broadcom, Cisco, Google, Linux Foundation, Microsoft, Nvidia, and Palo Alto. Financial firms like JP Morgan Chase have also joined, and Anthropic separately plans to grant access to over 40 organizations managing key software infrastructure.

Anthropic will first provide the preview version of its latest AI model, 'Mythos' (not yet released to the public), to participating organizations. In the 'CyberGym' evaluation, a benchmark for reproducing cybersecurity vulnerabilities, the Mythos preview recorded 83.1%, surpassing the previous top model 'Opus 4.6's' 66.6%. In the 'Humanity's Last Test' (HLE), which compiles problems at the PhD-level expert standard across various fields, it scored 56.8%, marking the first time an AI model has surpassed 50% in this evaluation.

Anthropic stated that it is providing this high-performance AI technology to major companies first, anticipating its potential misuse by hackers and aiming to give defenders an advantage. The ultimate goal, the company explained, is to safely deploy models at the level of 'Mythos' at scale.

Anthropic plans to offer up to $100 million in model usage rights to participating companies and donate a total of $4 million to open-source security organizations. The company has also been discussing 'Mythos's' cybersecurity capabilities with U.S. government officials, adding that it is prepared to cooperate with local, state, and federal governments to ensure the U.S. and its allies maintain a strong lead in AI technology.