Anthropic has officially declared the commencement of Project Glasswing, with its primary objective being the training of a cutting-edge, highly potent model named Claude Mythos Preview. This particular model was the subject of discussion in the recent cc source code leak incident and is a collaborative endeavor involving 12 institutions. However, for the time being, the model remains out of reach for the general public, with its overarching aim being to empower users to safely harness models of comparable capability in the long run. Presently, Anthropic intends to initially develop and rigorously validate safety mechanisms on the Claude Opus model. During its scanning of mainstream software, Claude Mythos Preview has unearthed thousands of zero-day vulnerabilities. In certain instances, it has demonstrated an ability to discover and exploit software vulnerabilities that surpasses that of the vast majority of individuals. Anthropic posits that AI technology can effectively lower attack costs, prompting it to proactively apply this technology to the defense sector. Project Glasswing will unfold in two distinct phases: initially, 12 founding partners and more than 40 organizations will gain access to the model. Anthropic has pledged to provide $100 million in usage credits and will clarify the pricing structure for commercial access once the research preview period concludes. Furthermore, Anthropic has made a generous donation of $4 million to bolster the maintenance of open-source software. Subsequent plans encompass information sharing and policy recommendations, wherein participants will engage in the exchange of information, publish research progress reports, and collaborate with security agencies to formulate practical recommendations.