Analysis: Why the Claude Mythos AI can't be released to the publicーNHK WORLD-JAPAN NEWS

Key Concepts

• Claude Mythos: A high-capability, general-purpose AI model by Anthropic specialized in cybersecurity.
• Zero-day flaws: Previously unknown software vulnerabilities that hackers can exploit before a patch is available.
• Attack Paths: A sequence of linked vulnerabilities that allow an attacker to move through a network to reach a target.
• GPT-5.5 Cyber: A rival cybersecurity-focused AI model developed by OpenAI.
• AI Sovereignty: The strategic concern regarding national dependence on foreign-developed AI for critical infrastructure.

---

1. Overview of Advanced Cybersecurity AI

G7 finance ministers recently convened in France to address emerging threats to the global financial system, with a primary focus on the risks posed by advanced artificial intelligence. The discussion centered on new, highly powerful AI models—specifically Anthropic’s "Claude Mythos"—which are currently restricted from public release due to their dual-use nature: they can either fortify or dismantle critical digital infrastructure.

2. Capabilities and Performance of Claude Mythos

Claude Mythos is designed as a general-purpose model with exceptional proficiency in cybersecurity.

• Technical Functionality: It excels at identifying "zero-day" vulnerabilities and mapping complex attack paths by connecting disparate system weaknesses.
• Performance Metrics: Testing conducted by the UK’s AI Security Institute revealed a 73% success rate on expert-level cybersecurity tasks.
• Efficiency: The model can simulate cyberattacks by chaining together sequences of actions across networks in seconds—tasks that would typically require human experts days or weeks to execute.

3. Risk Mitigation and Access Control

Because the same capabilities used to identify vulnerabilities can be used to exploit them, developers have implemented strict access controls:

• Restricted Access: Mythos is not available to the general public. Access is limited to select entities, including Apple, Google, and JPMorgan Chase.
• Competitive Landscape: OpenAI has developed a comparable system, GPT-5.5 Cyber, which is similarly restricted to trusted cybersecurity professionals. The UK AI Security Institute notes that GPT-5.5 Cyber performs on par with Mythos.

4. Geopolitical and Security Implications

The emergence of these models introduces significant national security risks:

• Critical Infrastructure: If exploited by bad actors, these models could disrupt electricity grids, water systems, communication networks, and global supply chains.
• The "Time Window": Experts estimate that major US labs are roughly 1–3 months ahead of competitors, while Chinese models are estimated to be 6–12 months behind. This provides a narrow window to patch existing vulnerabilities before these capabilities become more widespread.
• Global Influence: AI development is currently concentrated in the US and China. This concentration creates a geopolitical divide, where access to advanced AI becomes a tool for national security and global leverage.

5. Japan’s Strategic Position

Japan faces a significant challenge regarding its AI development:

• Investment Gap: While the Japanese government is increasing investment and fostering public-private partnerships, the scale of these efforts remains significantly smaller than that of leading global firms.
• Dependency Concerns: There is growing anxiety regarding Japan’s reliance on foreign technology for critical infrastructure, which could compromise national security if the supply or security of those models is disrupted.

6. Synthesis and Conclusion

The development of models like Claude Mythos marks a paradigm shift in AI capabilities. AI is moving beyond simple data processing to executing complex, independent tasks in high-stakes environments. The core takeaway is that the "frontier" of AI is no longer a future prospect; it is an immediate reality. The global financial and security landscape must now race to address the vulnerabilities these models expose before they can be weaponized by malicious actors. As noted by NHK deputy chief commentator Kawi, the future of AI-driven security is already here, and the ability to secure critical systems against these new tools is now a matter of national and global urgency.