China Considers Restricting Overseas Access to AI Models — Geopolitical Friction Expands Further
On July 7, Reuters reported that the Chinese government is considering restricting overseas access to AI models from major domestic AI companies such as Alibaba, ByteDance, and Z.ai. The Ministry of Commerce has reportedly been meeting with these companies one after another to discuss national security risks involved in providing AI models abroad.
The backdrop is the tightening of U.S. export controls on AI technology to China. In particular, some see this as a countermeasure against restrictions on the use of Anthropic's latest models (the Fable/Mythos families) in the Chinese market. It has also been suggested that violations could result in penalties under the national security law.
If these restrictions are enforced, companies and researchers that host Chinese open-source models (such as the Qwen series) on overseas servers or use them via API will be directly affected. For the open-source AI community, there are also concerns that the model distribution ecosystem could become fragmented.
OpenAI and Anthropic Offer Massive Computing Resources to Win Startups
According to WSJ information reported by The Decoder, OpenAI and Anthropic are intensifying their competition to acquire startups, wielding free computing credits as their weapon.
OpenAI is offering up to $2 million (approximately 300 million yen) in AI computing credits to Y Combinator startups. Anthropic, meanwhile, is offering credits worth $500,000 through the same program, but unlike OpenAI, it is emphasizing a condition that requires no equity in return.
In individual deals, there are also reports of cases where credits of up to $3 million (approximately 450 million yen) per startup were offered. Combined, the two companies could reach a total of $800 million (approximately 120 billion yen) per year, showing the extraordinary heat of the competition to lock in AI infrastructure. While these are attractive offers for startups, the risk of lock-in to specific platforms has also been pointed out.
AI Coding Assistants Quietly Expose Secrets — A Serious Security Risk
With the spread of AI coding tools (such as Claude Code and Cursor), a problem has emerged where developers unintentionally expose confidential information. According to an investigation by security researchers, a scan of 46,500 npm packages found that 428 of them contained AI tool configuration files such as `.claude/settings.local.json`, and 33 of those actually contained credentials such as API keys and tokens.
Furthermore, roughly one in 1,000 websites was found to have similar configuration files placed in publicly accessible locations. There is also a confirmed case where Apple accidentally included a CLAUDE.md file in its support app.
While AI coding tools dramatically improve development efficiency, local configuration files require the utmost care in handling. Organizational measures are urgently needed, such as adding these files to `.gitignore` or `.npmignore` and implementing automatic scanning before commits.
European Central Bank Steps Up Vigilance Against AI Cyberattacks
According to reports from the Financial Times and Bloomberg, the European Central Bank (ECB) and the European Systemic Risk Board (ESRB) have issued warnings about cyberattacks that exploit AI. They point out that the latest AI models could exploit vulnerabilities in financial institutions' IT systems within "minutes to hours."
The ECB has reportedly already asked major banks to draw up plans to respond to AI-driven cyber threats. This is based on the recognition that AI is dramatically streamlining the creation of phishing emails and the generation of malware code.
For not only financial institutions but all companies, cybersecurity measures in the AI era require responses that go beyond conventional frameworks, and building detection and response systems that leverage AI on the defensive side has become an urgent priority.
Australia's ABC Launches an AI Journalism Trial
The Australian Broadcasting Corporation (ABC) has announced that it will partner with Anthropic to launch a trial project for AI-powered journalism. According to a report by The Conversation, the initiative is expected to be limited mainly to format conversion of existing content, such as turning radio programs into articles.
ABC will hire AI-specialized staff and establish a workflow in which human editors ultimately review and approve the content. The positioning is not to replace journalists' work with AI, but to explore operational efficiency and new possibilities for expression.
This attempt at AI utilization by a public broadcaster will be an important precedent for the entire media industry. ABC's initiative is likely to offer many insights on how to leverage AI while maintaining the quality and credibility of news.
---
Bonus: Apollo Global's Economist Is Skeptical About AI Profit Forecasts
Torsten Sløk, Chief Economist at Apollo Global Management, has stated that it will take "much longer" than Wall Street expects for companies outside the tech industry to boost profits with AI. He points out that the ROI on AI investment is likely to be concentrated among big tech firms for the time being, adding fuel to the AI bubble debate.