Stay ahead in AI —
in just 5 minutes a day.

Practicality and safety move to the forefront

• The Fugaku NEXT project, driven by RIKEN, Fujitsu, and NVIDIA, redirected its efforts away from competing only on raw compute rankings and toward building a computer that’s actually usable in the AI era [1]. With the basic design completed, it moved into detailed design—and will incorporate GPUs to strengthen AI processing.
• This reflects a shift from chasing nothing but “numbers for speed” to thinking about how to balance scientific computing with AI workloads. Fujitsu also announced plans to sell domestically produced CPUs for AI data centers, strengthening the value of having Japan’s compute infrastructure at home [1].

AI coding tools’ vulnerabilities are laid bare

• The series of breaches involving OpenAI’s Codex, Anthropic’s Claude Code, and GitHub Copilot didn’t break what the AI “knows.” Instead, they exploited the handling of login information and permissions [2].
• For instance, techniques have been observed such as tampering with branch names to steal GitHub keys, or modifying configuration files to bypass restrictions [2].
• In other words, what’s dangerous isn’t only what the AI thinks, but which keys it has and how far it can go.

Safety feature upgrades—and expansion of what must be protected

• OpenAI introduced stronger protections for ChatGPT accounts and partnered with Yubico so users can connect key-based login flows rather than relying only on weaker methods [4].
• This shows that safety measures for AI services are expanding beyond preventing model misbehavior to include protecting accounts.

AI moves into public-sector work and everyday devices

• The Ministry of Land, Infrastructure, Transport and Tourism will, starting May 2026, gradually include guidance on how generative AI should be used in specifications for its direct oversight civil engineering operations [3]. This is a sign that AI is being integrated into how public work is carried out.
• Google is expanding Gemini to several million vehicles, enabling in-car capabilities such as navigation, conversation, and message summarization [8].
• Stripe is pushing toward a workflow where an AI-enabled wallet can handle actions like shopping and ticket purchases, with AI taking responsibility up to payment completion [7].

Don’t “trust AI all at once”—use it by separating roles

• First, it’s important not to treat AI as an all-purpose sidekick that can do anything. It’s strong in scenarios like brainstorming, drafting, organizing, and comparing—but for final decisions, especially where money or permissions are involved, it’s safer to keep the assumption that humans will have the last look [2][4].
• Next, before delegating anything to AI, decide what it is allowed to do. For example, pre-defining which documents it may view, which folders it may touch, and who it may reply to can reduce accidents while preserving convenience [2][5].
• And the more AI you bring in, the more you’ll need to change your own “using habits.” If you make it a routine to try quickly, question results, and review anything critical, you’re less likely to be thrown off by AI’s ongoing evolution.
• For general users, a good balance may be to use AI casually as a tool for reducing everyday hassles, while strongly protecting important areas like login and payments [4][7].
• For people using AI at work, it helps to view AI not as a “machine that speeds up work,” but as a drafting machine that produces material needing verification. That way, you can work with it without over-expecting it.