2026.06.23

The AI Compute Race and the Security Trap of Vibe Coding

The battle for open-source AI infrastructure has entered a new phase, with a wave of major compute contracts centering on SpaceX’s resources. Meanwhile, the shadow side of an era where anyone can be a developer is becoming visible: security risks in vibe-coded applications are starting to materialize. Domestically, AI implementation is steadily advancing in conservative industries like rail and manufacturing, broadening the reach of real-world AI. This edition tracks the industry across three axes: who controls compute, how to ensure the quality of AI-generated code, and where Japan’s AX strategies are heading.

Reflection AI Signs $150M/Month Compute Deal with SpaceX

Following Anthropic and Google, open-source AI startup Reflection AI has signed a major compute contract with SpaceX. Starting July 1, 2026, through 2029, the deal grants access to NVIDIA GB300 chips and surrounding hardware at SpaceX’s Colossus 2 data center near Memphis, Tennessee, at $150 million per month. The total contract value reaches up to $6.3 billion, though either party can terminate with 90 days’ notice after the first three months.

For scale: Anthropic holds a deal with SpaceX worth $1.25 billion per month, and Google’s runs $920 million per month — Reflection AI’s commitment is roughly one-tenth of those. Still, the company frames this as “one of the largest announced commitments to open AI infrastructure.”

Reflection AI was founded in 2024 by two former Google DeepMind researchers. It positions itself as an open-weight alternative to closed frontier labs like Anthropic and OpenAI — releasing trained model weights publicly, following a strategy similar to Meta’s Llama and Mistral. Securing this compute creates an unprecedented dynamic: the open-source camp now has the resources to train frontier-class models independently.

The timing is significant. With the U.S. government banning use of Anthropic’s closed models (Fable and Mythos), demand for open-weight alternatives in government procurement and security-sensitive sectors is surging. The debate between open and closed AI has become entangled with geopolitics and national security. Given that Colossus 2 runs the latest-generation GB300 chips, Reflection AI can now compete at the frontier training level in a single leap.

Source: TechCrunch - SpaceX inks compute deal with Reflection AI, an open-source AI lab

The rise of “vibe coding” — generating code quickly with AI — is coming at a cost: security risks are surfacing. A case reported by The Verge describes a project manager who used AI to build and immediately publish a website called “Boomberg” visualizing U.S. government spending. Months later, an SQL injection vulnerability was discovered that allowed attackers to read and write data they should never have accessed. The creator said: “It was a complete blind spot. An obvious oversight while I was learning this new technology. I’m sure others are making the same mistakes.”

This is far from an isolated incident. PocketOS founder Jer Crane posted on X that an AI coding agent deleted his entire production database. Serial entrepreneur Joe Procopio disclosed that a web app he built for personal use was unintentionally left publicly accessible.

From a technical standpoint, LLMs are optimized to generate code that works — but their ability to consistently apply security best practices is limited. SQL injection, XSS, and authentication/authorization misconfigurations are easily missed when the focus is on feature implementation. Non-engineer developers in particular tend to conflate “the code runs” with “the code is safe.”

The path forward is not to ban vibe coding, but to make security review a mandatory step in the workflow. This means automated SAST (static analysis) and DAST (dynamic analysis), penetration testing, and a culture of reviewing AI-generated code. From an entrepreneur’s perspective, the “AI-generated code + security audit” service market is forming right now. The larger the vibe-coding population grows, the greater the demand for services that close its security gaps.

Source: The Verge - Read this before you vibe-code another app

Seven Ways to Instruct Claude Code: Anthropic’s Official Breakdown

Anthropic has published an official blog post systematically covering how to give instructions to Claude Code. It explains seven methods — CLAUDE.md, Rules, Skills, Subagents, Hooks, Output styles, and Appending the system prompt — along with when and how to use each. ITmedia AI+ covered the overview.

The most fundamental layer, CLAUDE.md, is recommended to stay under 200 lines and be maintained by a single owner. Its primary role is to provide a codebase overview and serve as a table of contents pointing to other files — the key is not to overload it with detailed rules. File-specific constraints belong in Rules, and workflow procedures like deployments and reviews belong in Skills, creating a design that scales as the instruction set grows.

From an engineering perspective, Hooks and Subagents are worth particular attention. Hooks can execute shell commands at the moment a tool is called, enabling integration into CI/CD pipelines and automated enforcement of coding guidelines. Subagents enable orchestration of multiple agents, boosting capacity for large-scale parallel processing tasks.

In practice, Goodpatch — which deployed Claude Code company-wide — reported that 86% of employees with zero coding experience successfully shipped a deployment. The timing of this official guidance reflects the reality that Claude Code’s user base now spans everyone from non-engineers to senior developers. In an era where how you design your instructions directly determines tool performance, this kind of official framework carries substantial practical value.

Source: ITmedia AI+ - Claude Codeに指示を出す「7つの方法」と使い分け　公式が解説

Odakyu Launches AI-Powered Railway Crossing Safety System on June 24

Odakyu Electric Railway began live operation on June 24 of a system that detects people trapped at level crossings and automatically stops approaching trains. The initial rollout covers four crossings: “Shinnshinjuku No. 2,” “Shinnshinjuku No. 4,” and “Shinnshinjuku No. 5” between Shinnshinjuku and Sangūbashi stations, and “Mukōgaoka-Yūen No. 9” between Mukōgaoka-Yūen and Ikuta stations.

The mechanism is straightforward and effective. AI analyzes real-time footage from existing safety cameras. If a pedestrian, wheelchair user, cyclist, or motorcycle is detected remaining inside the crossing after the barrier has lowered, the system simultaneously sends a braking signal to the approaching train and activates a special warning light to alert the crew. Once the situation is resolved, the signal clears automatically. As a privacy measure, footage is not used for personal identification and is deleted within one year.

The system was co-developed by Meitetsu EI Engineers, Toyota Systems, and Toho Electric Industry, following validation trials that began in January 2023. The biggest challenge during more than three years of testing was maintaining detection accuracy across variable conditions such as nighttime and rain. For a safety-critical system, that level of caution before going live is appropriate — which is exactly why the transition to full operation carries such significance.

On the business side, Japan has tens of thousands of level crossings, with several dozen accidents still occurring each year. Given the social cost of each incident — service delays, liability, and reputational damage — demand from other railway operators to adopt this technology is virtually certain. The underlying technology also has clear potential for horizontal expansion into dangerous road intersections and factory floor safety management.

Source: ITmedia - 踏切に取り残された人をAIで検知→列車を自動停止　小田急が実運用

Daihatsu Automates Auto Parts Defect Inspection to 0.1mm Precision with AI

Daihatsu Motor has introduced an AI-powered visual inspection system on the aluminum machining line at its Shiga (Ryūō) Plant No. 1 facility. The system automatically inspects the interior of machined holes in transmission components for defects, replacing a process that previously depended on the eyes and judgment of skilled inspectors.

The numbers convey the difficulty: the system must detect differences on the order of 0.1mm that affect product quality, while defect types and locations vary widely — making it impossible to achieve stable accuracy with conventional rule-based image recognition. Inspectors faced significant eye strain, and the process had long been highly dependent on individual experience and intuition.

The system was co-developed with VRAIN Solution (Chūō, Tokyo), combining field-accumulated expertise with AI and image recognition technology. The result achieves both stable accuracy and reduced operator burden. Horizontal expansion to other aluminum products with similar structures is already being considered, and both companies have filed a joint patent.

In the broader context of manufacturing AI, this case is one answer to the long-standing challenge of digitizing skilled workers’ tacit knowledge. Quality inspection of automotive parts is labor-intensive and faces growing recruitment difficulties. Manufacturers like VRAIN Solution — specialized in industrial AI — are well-positioned for significant growth by expanding to Tier 1 and Tier 2 suppliers. As labor shortages in Japanese manufacturing deepen, demand for solutions that replace “expert human eyes” with AI will continue to grow.

Source: ITmedia AI+ - ダイハツ、自動車部品のキズ検査をAIで自動化　“人の目と感性”を代替

Hitachi CEO: “AI Is Not a Threat, but a Growth Opportunity” — A Major Bet on Physical AI

The AX (AI Transformation) strategy presented by Hitachi President and CEO Toshiaki Tokunaga at “Hitachi Investor Day 2026” on June 10 has drawn significant attention. His remarks to investors and analysts offer a clear window into how one of Japan’s most established industrial conglomerates intends to navigate the AI era.

The core of Tokunaga’s argument is “Physical AI.” In a world where AI moves beyond cyberspace to directly influence the physical world — controlling energy infrastructure, optimizing mobility, automating production lines — AI malfunctions can directly cause equipment shutdowns or quality failures. This is precisely why companies capable of deeply integrating IT (information technology), OT (operational technology), and Products (physical equipment) are uniquely positioned to deliver safe implementations. Hitachi, he argues, is one of the few companies that holds all three.

On the digital side, he noted that while AI is increasingly replacing routine tasks like coding, demand is surging for “AI-readying” (modernization) of the more than 15,000 IT systems currently in operation. This is Hitachi’s core battleground as a systems integrator — and an arena where competition with Accenture and NTT Data is intensifying.

“AI is not a threat but a growth opportunity” carries some weight as positioning for shareholders, but Hitachi’s actual business portfolio — spanning energy, rail, and IT services — genuinely places it at the frontier of Physical AI deployment. The company’s conglomerate diversity has sometimes drawn criticism from advocates of focus and consolidation, but in the context of AI integration, that very diversity becomes a competitive advantage. That irony is worth noting.

Source: ITmedia AI+ - 日立はAX事業にどう臨む？　徳永CEOの話から「成長につなげるための勘所」を探る

Summary

If this week had to be captured in a single phrase, it would be “the tension between AI democratization and quality assurance.” Reflection AI’s SpaceX deal signals that the open-weight camp is establishing a position where it can compete head-on with closed models on compute — the most physical dimension of the AI race. Meanwhile, security vulnerabilities in vibe-coded apps represent an unavoidable consequence of AI lowering the barriers to software development. Inside Japan, from railway crossings to automotive parts inspection, quiet, practical implementations of “solving hard real-world problems with AI” are accumulating — a steady societal integration of AI that runs on a separate track from the headline-grabbing advances of frontier models. The direction of the industry as a whole suggests that the maturation phase — moving from “anyone can use AI” to “anyone can use AI safely and correctly” — is only just beginning.