Too dangerous or just too expensive? The real reason Anthropic is hiding Mythos (kingy.ai)

Anthropic recently initiated a gated rollout of its latest AI model, Claude Mythos Preview, under Project Glasswing, which limits access to around 40 vetted organizations, primarily in cybersecurity and critical infrastructure. Unlike typical AI releases, Anthropic refrained from broad marketing efforts, citing security concerns over Mythos's unprecedented capability to autonomously discover and exploit zero-day vulnerabilities at scale. This cautious approach aims to buy time for defenders to enhance safeguards before such potent offensive tools become widely accessible, according to the company's documentation. However, skepticism surrounds this narrative, with some observers suggesting that underlying compute costs and infrastructure constraints may also be significant factors driving Mythos's restricted access. Reports reveal that Anthropic is under pressure to secure additional computing resources, having announced partnerships with Google and Broadcom to expand capabilities. The economics of running a model like Mythos, particularly given its extensive context window (up to 1 million tokens) and high operational costs, complicates its release further. This dual narrative of security and economic feasibility raises critical questions about Anthropic's strategy and the broader implications for AI model accessibility and safety in the industry.