Anthropic co-founder Jack Clark made his first visit to Japan and said AI tools may be used in cyberattacks occurring around the world. He spoke at a media briefing held by Anthropic’s Japan subsidiary on October 14. Clark said hackers use available tools just as companies do, but told reporters he had not confirmed any case of AI carrying out an attack autonomously.

Anthropic said it is addressing misuse through safeguards and Project Glasswing, which gives companies priority access to its latest AI models. Clark said Anthropic’s own measures cannot improve the situation on their own because other companies’ models and open-weight models are also available. He said the company is spending time on policy discussions aimed at “common-sense traffic rules” that apply equally to all providers and create a safe environment for everyone.

Clark also discussed recursive self-improvement (AI improving itself to raise its capabilities). Referring to his newsletter statement that this could become possible as early as 2027, he said evidence had emerged this year that AI systems were beginning to show creativity on scientific problems ranging from mathematics to biology. He said AI research is a form of science and that people should allow for the possibility that AI-driven science will begin advancing at a very fast pace next year.

Hidetoshi Tojo, president and representative director of Anthropic’s Japan subsidiary, also spoke at the briefing. On cyberattacks, he said the emergence of Mythos in the spring made them foreseeable, but added that measures could still be taken and urged people to use AI in responding. The remarks were presented as the speakers’ views; Clark described AI’s use in attacks as a possibility, not a confirmed fact.

Anthropic co-founder Jack Clark made his first visit to Japan and said AI tools may be used in cyberattacks occurring around the world. He spoke at a media briefing held by Anthropic’s Japan subsidiary on October 14. Clark said hackers use available tools just as companies do, but told reporters he had not confirmed any case of AI carrying out an attack autonomously.

Anthropic said it is addressing misuse through safeguards and Project Glasswing, which gives companies priority access to its latest AI models. Clark said Anthropic’s own measures cannot improve the situation on their own because other companies’ models and open-weight models are also available. He said the company is spending time on policy discussions aimed at “common-sense traffic rules” that apply equally to all providers and create a safe environment for everyone.

Clark also discussed recursive self-improvement (AI improving itself to raise its capabilities). Referring to his newsletter statement that this could become possible as early as 2027, he said evidence had emerged this year that AI systems were beginning to show creativity on scientific problems ranging from mathematics to biology. He said AI research is a form of science and that people should allow for the possibility that AI-driven science will begin advancing at a very fast pace next year.

Hidetoshi Tojo, president and representative director of Anthropic’s Japan subsidiary, also spoke at the briefing. On cyberattacks, he said the emergence of Mythos in the spring made them foreseeable, but added that measures could still be taken and urged people to use AI in responding. The remarks were presented as the speakers’ views; Clark described AI’s use in attacks as a possibility, not a confirmed fact.