Why Is Cutting-Edge AI Not Made Available to the General Public?

◆The generative AI model “Claude Mythos Preview,” announced by the US-based company Anthropic in April 2026, is considered to possess advanced capabilities in the cyber domain and is currently available only to select institutions and companies. This decision to limit access serves as an early example indicating that we are approaching a stage where existing systems and response capabilities are no longer sufficient to fully manage the social impact of AI.

◆These developments suggest that the way AI models are made available may split into a two-tier system: AI made widely available to the general public, as has been the case so far, and AI with restrictions on users and applications. While this represents a rational risk management strategy for AI model developers in high-risk domains, it may also create disparities in capabilities and competitiveness among users, depending on whether they have access to these models.

◆However, it is not certain that AI models will continue to be made available in this way going forward. Future developments are likely to be shaped by two key axes: “the timing of the widespread adoption of AI models on a par with frontier models” and “the stringency of access governance.” Based on these axes, the structure of AI model provision can be categorized into three scenarios: (1) the expansion of controlled-access AI (cases where governance functions effectively), (2) the fragmentation of governance (cases where an unstable two-tier structure emerges), and (3) the erosion of governance (cases where AI becomes broadly available in practice).

◆Currently, we are in a situation corresponding to Scenario (1) (the expansion of controlled-access AI) described above; however, the widespread adoption of AI models comparable to frontier models and changes in access governance could lead to a transition to other scenarios. In particular, if we move toward Scenario 3 (the erosion of governance), there will be little time to implement measures against the risk of misuse. Therefore, companies must establish rapid response systems based on the assumption that damage will occur, while governments must take diplomatic measures to secure access to AI models and consider sovereign AI strategies.