Weekly AI news: Agents, Workflows and Security

This week’s AI news points to a practical question: how should companies connect capable models to everyday work without losing oversight? Drawing on my X posts from October 7–11, 2026, I review model selection, workplace tools and security, then explain how I would approach adoption in Japan. I also cover my own work on presentation materials and Digirise’s latest company meeting.

Model selection: Gemini Agent, Grok bot and Claude credits

On October 9, I covered Google’s new Gemini Agent service, which includes Claude Opus 5.5 and Sonnet 5.5. It automatically selects models according to the task, with additional third-party models planned. The scope matters: this announcement concerns the new service, not the consumer Gemini app, where support has not been announced.

My October 8 post covered a related change in direction for Grok bot. Elon Musk outlined an approach that combines external APIs rather than relying solely on xAI. The selection criterion was the highest probability of producing the best result for a task. Claude Opus 5.5 was named as a candidate, alongside plans to use MidJourney for images and Suno for music. A named candidate should not be confused with an already available integration.

I see the useful question shifting from which model a company prefers to how it assigns work and checks results. Automatic selection does not remove the need to define a successful outcome.

I also covered monthly API credits for Claude subscriptions:

• Max 5x: $100 per month
• Max 20x: $200 per month
• Team: up to $500 per month, pooled across the organization

The credits cover all models, including Haiku 5.5, and support users’ own code or third-party harnesses. I would start with a clearly defined internal prototype rather than assume the allowance will cover any workload.

Workplace AI: editing files, recording meetings and generating interfaces

Claude’s integration with Google Docs, Sheets and Slides puts it inside the file being used. The sidebar can read and directly edit that file. My post covered formatting-preserving revisions in Docs; formulas, pivots, charts and new tabs in Sheets; and additional slides using an existing deck’s layout and theme.

The default “Ask before edits” mode is important to my adoption judgment. I would use those approval points to understand the changes before expanding what I delegate.

ChatGPT’s Meetings plugin records meetings and saves summaries and next actions to Space afterward. At the time of my post, it was a beta feature in the macOS desktop app for Pro and Business, with Enterprise coming soon. Notes can be saved privately or shared with a team. Users can then request project-plan updates or reply drafts.

For me, the value is not simply producing a transcript or summary. It is connecting a meeting to subsequent work. I would review the summary and actions before asking AI to update a plan.

On October 8, I also covered Intelligent UI, described in my post as beginning rollout alongside GPT-6. It generates interactive interfaces for responses, with examples including bicycle diagrams, dinner-plan cards, wardrobe suggestions and receipt-based bill splitting. I would test it where a visual presentation makes the result easier to inspect, rather than assume every answer needs an interface.

Practical agents: presentations, messages and a ring-shaped interface

In my own work, I shared how Claude Code can create seminar materials exceeding 100 pages, with animations and PowerPoint-like editing. I also posted an update on our internal presentation-creation tool.

The distinction I would preserve is between generating many pages and communicating a clear argument. I would define the audience, conclusion and structure first, then review the resulting content and sequence. Faster creation should leave more room for that review.

Another post described using Grok bot’s own browser to log into LINE and Messenger and check messages every 30 minutes. The approach brings message checking together without using APIs and can also send automated replies.

I would treat reading and sending as separate decisions. Reducing missed messages is not the same responsibility as allowing an agent to speak on someone’s behalf. For a company trial, I would define what it may inspect and what still requires approval before sending.

Natura’s Interface offers a different entry point. The ring-shaped device sits at the base of the index finger; the user presses its single button with a thumb and speaks. It connects to agents including Claude Code, Codex and OpenClaw, with responses delivered through earphones or the iPhone lock screen. My post listed an early-purchase price of $99 and shipping planned for January.

I am interested in giving instructions without opening a phone. For adoption, however, I would first identify a task where that interaction is genuinely useful.

Security: attacker claims, management responsibility and OSS scanning

On October 7, I covered ransomware group Qilin’s response to a JNN interview. The group, which attacked Asahi Group and others, sharply criticized Japan’s security and suggested attacks on Japanese businesses and government bodies could increase. Those are statements from an attacker, not an objective assessment of the entire country.

My post also covered the group’s claim of working with more than 300 people worldwide and more than 1,000 confirmed incidents in 2025. In IPA’s Information Security 10 Major Threats 2026, ransomware, supply-chain attacks and AI-related cyber risks ranked first, second and third respectively. As AI lowers attackers’ costs, I regard security as a management issue.

Japan’s Personal Information Protection Commission also revised its WARNING notice on October 7. Based on actual breach reports, it groups incidents into nine categories and adds API abuse. It also addresses inadequate oversight of group companies, overseas locations and cloud services, and notes possible violations of Articles 23–25 where measures are insufficient.

On the defensive side, I covered Anthropic’s OSS Scanner, a free service requiring an application. It uses models including Mythos for periodic scans and delivers reproduction code and patches.

The figures need careful separation: 29,000 vulnerability candidates were found over the preceding six months, while about 6,000 had been reviewed by people. In a validation exercise involving 97 serious candidates, 85 were genuine—88%—and one was a false positive. Candidate discovery is not the same as confirmation, and that result should not be generalized to every finding.

I would define who reviews findings and authorizes fixes before treating scanning as an operational process.

Building an AI business: Cursor and Digirise’s next stage

On October 11, I covered Musk’s explanation of SpaceX’s acquisition of Cursor. He named Anthropic, OpenAI and Cursor as the three companies with sufficient coding-training data and described data, talent, compute and enterprise sales as requirements for success.

My post also included Cursor’s sales team serving thousands of companies, the $60 billion acquisition price and the acquisition of xAI’s Colossus compute infrastructure. I treat the judgments about those companies as Musk’s explanation. The question I take from it is how technical capability becomes something businesses can actually adopt.

At Digirise, we held our company-wide meeting for our fifth fiscal period on October 7. We were three years and three months from founding. Revenue in the previous period was five times the preceding period, and the team had grown to around 50, including full-time contractors.

For the new period, I set out our direction as an “AI trading company,” adding physical AI and security businesses to education, consulting and development. With GMO’s help, we purchased a G1 humanoid. At the time of the post, delivery was scheduled for Friday; these posts do not contain post-delivery test results.

I want to examine the technology hands-on. In management, I also emphasized real-time visibility into data and using AI to accelerate decisions and development. Our goals include going public and building a lean, high-profit organization.

What Japanese adopters and overseas vendors should take from this

For companies in Japan, I would put availability and oversight ahead of product names. Gemini Agent’s scope, Meetings’ supported environment and Claude’s editing approvals are practical conditions for a trial, not minor details.

For overseas vendors entering Japan, my recommendation is to explain a specific workflow. Show which files the product uses, what it changes and where a person checks the result. I believe that is more useful for an adoption decision than a broad claim about model capability. Explain the operational boundaries alongside the convenience.

How I would apply this inside a company

I would choose a bounded task in presentation creation, post-meeting follow-up or message review. I would define the information the tool can use and the actions requiring approval, inspect the outputs, and only then expand its responsibilities. The aim is progress in the work—not simply adding more AI products.