A growing number of advanced AI systems are now being deliberately restricted by their creators due to safety concerns, marking a major shift in how cutting-edge technology is deployed. According to a recent Time report, models such as Claude Mythos by Anthropic and GPT-Rosalind by OpenAI are not being made publicly available. Instead, they are released only to select, “trusted” users through controlled access programs, reflecting fears about the risks these powerful systems could pose if widely distributed.
The main concern lies in the dual-use nature of these AI models. Systems designed for beneficial purposes—such as cybersecurity research or biological discovery—can also be misused. For example, tools that help identify software vulnerabilities could be used for cyberattacks, while AI models trained for life sciences could potentially assist in designing harmful biological agents. Experts note that “cyber defense and cyber offense look very similar,” highlighting how difficult it is to separate beneficial and dangerous uses.
This shift has sparked a broader debate about who should control access to such powerful technologies. Currently, private companies are making key decisions about which users qualify as “trusted” and how these systems are deployed. However, policymakers and researchers argue that governments should play a stronger role, given the potential societal and national security risks. The issue raises deeper questions about accountability, oversight, and whether corporate self-regulation is sufficient for technologies with such far-reaching consequences.
Looking ahead, experts warn that restricting access may only be a temporary solution. Open-source AI models are rapidly improving and could soon match the capabilities of restricted systems, potentially making these powerful tools widely accessible regardless of corporate controls. This creates a complex balancing act—ensuring that beneficial uses of AI continue to advance while preventing misuse—suggesting that stronger global governance and coordinated safeguards may become essential in the near future.
