【Claude Mythos級オープンモデル拡散で始まる悪夢】チームみらい・安野貴博「従来型サイバー防衛ではもう無理」／「1～2年で発生」今井翔太が懸念する「GPT暴走」以上の事件【AI QUEST】

Summary

AI researcher Shota Imai and Team Mirai leader Takahiro Anno discuss how Japan should respond to rapidly advancing AI, focusing on cyber risk, international competition, government adoption, and employment.

Cybersecurity is becoming more urgent

Anno says that after Anthropic announced Claude Mythos Preview, he expected the gap between frontier models and open-weight models to close quickly. The speakers cite reports suggesting that an open-weight model may already have cyber capabilities approaching that level. They warn that current attacks may not yet reflect the strongest available models, and that more capable tools could spread soon.

Anno argues that conventional defenses, such as periodically checking for weaknesses, will no longer suffice. Organizations will need to continuously probe their own systems using powerful models.

Prioritize AI risks by likelihood and urgency

Anno distinguishes three broad dangers: AI eventually causing human extinction; a catastrophic confrontation between major powers using AI; and terrorists or other malicious actors using AI. He considers the latter two more immediate than a science-fiction-style AI rebellion.

He also points to biosecurity risks, including AI-assisted design of DNA or RNA sequences, and argues that practical safeguards in synthesis services should be implemented now.

A pause in the AI race seems difficult

The speakers say that companies and countries have strong incentives to keep advancing, especially amid US–China competition. Imai worries that a serious incident involving AI agents could occur within a year or two and abruptly change public attitudes.

They also discuss how competition for computing resources can shape research priorities, even within major AI companies.

A proposed strategy for Japan

Anno proposes a three-part strategy:

1. Use Japan’s strengths in semiconductor manufacturing to gain leverage.
2. Keep domestic AI capabilities close enough to the frontier to benefit from future shifts in technology.
3. Collaborate with other mid-sized powers—such as France, Canada, Australia, Germany, and South Korea—on shared AI infrastructure and research.

He presents this as a difficult but more realistic path than trying to match the US or China alone.

France as an example of pragmatic AI policy

Anno says French public institutions make substantial use of open-weight models, including Chinese models, rather than relying exclusively on American services. The discussion emphasizes that not every task requires the most powerful—and expensive—frontier models. Governments can use smaller models where suitable while building systems to assess the safety of the models they deploy.

Government adoption requires AI access and a realistic budget

Anno argues that Japan should allocate roughly ¥100 billion for government AI use, including the ongoing costs of agent-based tools. AI could help civil servants process documents, compare precedents, and develop policy, freeing them to focus on higher-value work.

The speakers also argue that domestic model projects need clear priorities and concentrated decision-making authority. Slow consensus processes may hinder ambitious research.

Employment effects may appear first among junior workers

The speakers cite overseas research indicating declines in some junior-level roles, including in call centers, software engineering, and consulting. They warn that even if expert jobs remain, the entry-level work that helps people become experts may disappear. Imai argues that society may need new ways to develop talent outside the usual workplace training model.

Detect disruption early and respond quickly

Anno calls for more timely labor-market indicators, adaptable retraining programs, and support for workers who cannot readily change jobs. He also says governments should begin discussing broader social protections, potentially including basic-income-like measures, before large-scale displacement occurs.

AI displacement can coexist with labor shortages

Japan may continue to have understaffed industries even as some occupations lose jobs. The speakers suggest that social expectations, status, and reluctance to move between sectors can impede labor mobility.

They also discuss the possibility that work valued chiefly for practical productivity may be especially exposed to automation, while some creative or competitive activities may retain a distinct human appeal.

Presenters and contributors

Shota Imai and Takahiro Anno.