Japan reportedly asks Anthropic and OpenAI for wider access to AI models that find security flaws

Japan's government is finalising an emergency cybersecurity package after a series of attacks on domestic companies, and according to a Nikkei report it intends to lean more heavily on American AI developers. On October 30 the newspaper reported that Tokyo has asked Anthropic and OpenAI for greater access to advanced models capable of spotting security vulnerabilities, and that it will encourage operators of critical infrastructure to use them. Nikkei's article is paywalled; outlets relaying it, including FourWeekMBA and PANews, say the readable text does not name specific models or indicate whether either company has agreed. As of the most recent reports reviewed, the package itself had not been formally announced.

What the government ordered

The push follows the first ministerial meeting on cybersecurity, held at the Prime Minister's Office on October 26. Prime Minister Sanae Takaichi told ministers the situation had become serious and asked for a package of emergency measures to be compiled within about a week, according to AFP and Kumamoto Asahi Broadcasting. Among her instructions, she asked officials to use AI to simulate attacks that expose weaknesses in systems, and to bring forward security checks of government networks. Other steps include more consultation support from police and other agencies for affected firms, faster sharing of attack information, and work with allies to identify attackers. The Yomiuri Shimbun reported that the government is also weighing active cyber defense, under which authorities could neutralise attacking servers.

Why AI is central

Officials see AI on both sides of the problem. The National Cybersecurity Office presented an analysis indicating that attackers are likely using AI, the Yomiuri reported. AFP noted that GMO Internet Group, itself a target, said a hacker had demanded 50 million dollars, and that its security unit linked several recent intrusions to a Chinese AI security-testing tool called Artex.

Japan already has some access to such defensive models. SmartScope recalls that OpenAI said in May it would provide GPT-5.5-Cyber to the Japanese government and approved companies. On October 13, according to PANews citing Nikkei, cybersecurity minister Shunji Furukawa met Anthropic co-founder Jack Clark, and the two sides agreed that Anthropic would work with a government-affiliated AI safety institute on how AI could be used in attacks. What the expanded access would cover, and on what terms, has not been disclosed.