Friday, August 21, 2026 Trend Press · Cloudflare Pages

The Trend Tribune

"All the trends that are fit to read" Morning Edition Free of Charge
TODAY'S LEAD STORY

"All S7s" are targeted: AI is changing the economics of attacks and a warning about water infrastructure.

This article explains the warning issued jointly by CISA, FBI, and NSA on August 20th regarding attacks targeting all Siemens S7 series PLCs, within the context of the changing economics of AI-driven cyberattacks. It summarizes the actual damage to water facilities in more than 12 states across the US, the feedback loop structure of agent-based attacks involving multiple AI systems working together, the reality of the difficulty and degree of automation in attribution as shown in the International AI Safety Report 2026, and counter-trends such as the use of AI on the defense side.

By tech_tanaka · August 21, 2026
"All S7s" are targeted: AI is changing the economics of attacks and a warning about water infrastructure.
(Photo: illustrative)

"All S7s" Targeted: AI-Driven Changes the Economics of Attacks and a Warning for Water Infrastructure

On August 20th, multiple government agencies, including the US Cybersecurity and Infrastructure Security Agency (CISA), the FBI, and the NSA, issued a joint warning. The warning stated that attackers were targeting Siemens' S7 series industrial control devices, and that the damage had already affected water facilities in more than 12 states across the US. What's noteworthy for engineers is not the scale of the damage itself, but the fact that this warning was issued within the context of "AI changing the economics of attacks."

An Unusually Broad Warning Scope: "All S7s"

What's distinctive about this warning is its broad scope. CISA explicitly states that it targets "all" Siemens S7 series PLCs (Programmable Logic Controllers, specialized computers that automatically control machinery in factories and facilities), used to control automated physical processes in a wide range of sectors, including water, energy, manufacturing, and agriculture. This is a broader alert than usual, indicating that the entire product family is being targeted, rather than a warning focused on a specific model number or vulnerability.

Confirmed damages include decreased water pressure, flooding, recommendations for boiling water for disinfection, and cases where users were forced to switch from automated to manual control. These demonstrate that cyberattacks are not merely abstract events in the world of data, but are actually having physical consequences that affect people's daily lives.

Agent-Based AI Changing the "Economics of Attacks"

This news is more significant than a single vulnerability report because it is part of a larger trend: agent-based AI is changing the very nature of cyberattacks.

Traditional cyberattacks, while using scripts and automated vulnerability scanners, have historically relied on human operators for critical decision-making. However, with the advent of agent-based AI (AI systems capable of autonomously performing multiple tasks), a "feedback loop" is theoretically becoming possible, where multiple AI systems share tasks, analyze each other's findings, and continue attacks with limited human intervention. If this actually works, relatively small groups could launch simultaneous attacks against a large number of targets at machine-like speeds, potentially fundamentally changing the very "economics" of cybercrime and state-sponsored espionage.

The Issue of "Attribution Difficulty": A Matter to Handle Calmly

However, caution is necessary when evaluating such discourse on "autonomous AI attacks." Technically identifying who carried out the attack—"attribution"—remains an extremely difficult task, and the claim that "this was an AI-driven attack" itself requires careful scrutiny.

In fact, the "International AI Safety Report 2026," compiled by an international group of experts, offers a calm assessment of this point. In November 2025, an AI development company reported a case where a threat actor used its model to automate 80-90% of the intrusion activity, with human involvement limited to critical decision-making points. The researchers have also demonstrated that AI systems can autonomously search for vulnerabilities in computer networks in laboratory environments. However, the report also notes that there have been no reported instances of general-purpose AI systems executing fully autonomous end-to-end cyberattacks in the real world.

The Counter-Movement: "Defenders Also Use AI"

In parallel with the CISA and other warnings, it's important to note that not only attackers but also defenders are beginning to utilize AI. This means a new form of competition is emerging, where both attackers and defenders are increasingly facing off against each other with automated systems.

The proliferation of AI-powered monitoring and detection systems to protect critical infrastructure, such as OpenAI's defender model "GPT-5.6-Cyber," which we previously discussed, and the issues raised in this warning, can be seen as part of this trend of "automated defense."

What Engineers Should Consider

This joint government warning is significant because it links the abstract discussion of "AI risk" to concrete damage that is actually impacting people's lives. CISA is urging potentially affected businesses to patch affected software, isolate controllers exposed to the internet, secure remote access, and refresh outdated credentials.

For organizations operating industrial control systems, this warning should be viewed not only as preparation for the future risk of "automated AI attacks," but also as an opportunity to address the immediate challenge of "a lack of basic security hygiene" that is already underway. Before discussing the sophistication of attacks, the most basic and effective first step is likely to be ensuring that their control systems are not vulnerable to the internet.

References: techcrunch.com / securityboulevard.com / techstartups.com / arxiv.org
サイバーセキュリティ重要インフラ産業制御システムAIエージェントAI/ML

Creating a "Matrix for Robots": A Former NVIDIA Researcher Takes on the Biggest Bottleneck in Physical AI

By robo_takahashi

On August 19, Sanja Fiedler, former Deputy Head of AI Research at NVIDIA, announced the launch of Veeda AI, a global modeling startup, and its seed funding of over $90 million. This article explains the bottleneck in physical learning where robot hardware does not scale like computing resources, the founding team's track record at NVIDIA Spatial Intelligence Lab, its positioning among leading global modeling companies such as AMI Labs, World Labs, and Waabi, and its four-city network in Toronto, Zurich, Singapore, and Mountain View.

Creating a "Matrix for Robots"—A Former NVIDIA Researcher Tackles the Biggest Bottleneck in Physical AI

On August 19th, Sanja Fiddler, former Deputy Head of AI Research at NVIDIA, announced the launch of her new company, Veeda AI, and the raising of over $90 million in seed funding. In a LinkedIn post, Fiddler described what her company is developing as a "matrix for physical AI." As a software engineer, I want to examine the technical challenges this company is trying to tackle.

Why Can't Robots Learn Through "Real-World Trial and Error"?

Fiddler's concern is clear. In her LinkedIn post, she points out that the very idea that robots can learn through real-world trial and error is unrealistic. The reason is simple: robot hardware doesn't scale like computing resources.

For learning large-scale language models, adding additional GPUs allows thousands of trials to run simultaneously in parallel. However, when it comes to having robots learn through trial and error in the real world, the sheer number of physical robot components—the "hardware" itself—limits the learning speed. Furthermore, repeated failures in the real world lead to component wear and tear, and sometimes even safety risks. This "physical learning bottleneck" is the biggest structural challenge facing physical AI (the field that applies AI to actions in the physical world).

The "World Model" Approach

Veeda is working on an approach to solve this problem using a technology called the "world model." A world model refers to an AI model that learns physical laws and the behavior of objects, and can generate simulated environments and situations. The idea is that if robots can repeatedly learn through trial and error in an infinitely scalable simulation environment, learning can be accelerated on a large scale without being limited by real-world hardware constraints.

The company aims to build an infinitely scalable simulation environment where robotics developers can train embodied AI agents through repeated interactions. In Fiddler's words, the company's sole mission is to build a "simulated reality for physical AI."

Technical backing demonstrated by their NVIDIA track record

Important to evaluating the persuasiveness of this approach is the background of Fiddler and his co-founders, who led research in this very area at NVIDIA. Fiddler led NVIDIA's Toronto-based Spatial Intelligence Lab for eight years, where he was involved in developing software that simulates how self-driving cars and humanoid AI agents interact with the real world. Co-founder Huang Lin (Chief Scientist) also worked in the same lab, and the other co-founder, Zhan Gojičić (CTO), was a research director at NVIDIA based in Zurich.

In other words, Veeda can be understood not as a startup building technology from scratch, but as a company established as a platform for a team with some of the industry's most advanced knowledge in world model research to further pursue that technology.

Differences in Positioning from "World Labs" and "AMI Labs"

The field of global models is not something Veeda is new to. "AMI Labs," founded by former Meta AI Chief Scientist Yann LeCun, raised $1.03 billion in March, and "World Labs," founded by Fei-Fei Li, also raised $230 million. Waabi, another Toronto-based autonomous driving company, is also developing global models as part of its autonomous driving technology.

Compared to these pioneering companies, Veeda's $90 million seed funding may seem small in absolute terms. However, it is noteworthy that this is one of the largest seed funding rounds ever for a Canadian company, achieved in just three months since its founding. The scale of this funding highlights the capital-intensive nature of the global model field, requiring both high-quality research talent and sufficient computing resources.

Things to Note from a Software Engineer's Perspective

Veeda's concept of a "matrix for robots" addresses one of the most urgently needed technological pieces in the development of humanoid robots and autonomous vehicles. While many of the humanoid companies we've covered so far compete on their operational track record with actual robots, the "simulation platform" layer, which streamlines the learning process itself, is relatively understated but represents a fundamental technological area that can resolve bottlenecks across the entire industry.

The fact that Veeda's offices are planned to be located in four locations—Toronto, Zurich, Singapore, and Mountain View—reflects the global dispersion of talent in this field. We look forward to seeing how many robot development companies actually adopt this global model and how much it can improve the learning efficiency of actual robots, as concrete implementation examples emerge.

References: thelogic.co / siliconangle.com / betakit.com / www.theglobeandmail.com
フィジカルAI世界モデル資金調達ロボティクスNVIDIA

From "1 in 1000" to "1 in 2"—Real-world data reveals the two-year trajectory of AI coding.

By thesis_yamamoto

This article analyzes the "How teams build" report, a real-world usage data report published by the project management tool company Linear. It explains, from a researcher's perspective, the reasons behind the surge in AI-generated issues from less than 1 in 1,000 to approximately half in two years, the tripling of weekly pull requests (PRs) for teams using coding agents from 21 to 65, the paradox that while the AI ​​adoption rate among CEOs increased from 9% to 36%, overall development time actually increased, and the interpretation of adjustment costs as "AI is an add-on, not a replacement."

From "1 in 1,000" to "1 in 2"—Real-World Data Reveals the Two-Year Trajectory of AI Coding

Linear, a provider of project management tools, has released a report titled "How teams build," which compiles data from its own products. Based on millions of actual usage data points, it quantitatively demonstrates the changes brought about by AI in software development. As a journalist with a background in AI research, I want to carefully analyze the changes revealed by real-world data, rather than focusing on flashy announcements or individual technical demonstrations.

From "Less than 1 in 1,000" to "Approximately Half"

The most symbolic figure shown in this report is the change in the percentage of issues (records representing work items or tasks) created by AI on Linear. Two years ago, issues created by AI accounted for less than 1 in 1,000. Now, AI creates approximately half of almost all issues created on Linear. Furthermore, this report predicts that "at the current pace, the amount of code created by AI will soon surpass the amount created by humans and other integrated tools combined."

This rapid change in numbers suggests a structural shift in the development workflow itself, which cannot be simply attributed to the "popularization of AI tools."

Coding Agent Usage Increases from 21 to 65 Per Week

Figures showing a more direct impact on coding work have also been released. Teams using coding agents saw a more than threefold increase in the number of pull requests (code change proposals) per week, from 21 to 65. In contrast, teams not using agents saw only a modest increase, from 8 to 10.

This contrast indicates that the introduction of AI coding agents goes beyond simply "making the work easier" and structurally boosts the amount of code changes that teams can generate.

"Adoption Rate Doubles"—Particularly Significant Growth Among Management

Interesting data is also presented regarding the adoption rate of AI features. During the first half of 2026, the adoption rate of AI functions more than doubled across all job functions. Particularly noteworthy is the adoption rate among CEOs in companies with 201 or more employees, which quadrupled from 9% to 36%.

The fact that management is actively beginning to use AI tools themselves is a remarkable sign that AI adoption is no longer limited to engineers on the front lines, but is permeating decision-making levels throughout the entire organization.

Paradoxical Finding: "Development Time Still Increases"

The most interesting finding from a researcher's perspective in this report is a seemingly counterintuitive finding. From June 2025 to June 2026, the time spent on issue creation, triage, and commenting increased across almost all job functions. In engineering roles in particular, the time spent on creation and triage alone increased by approximately 17%. Founders, in particular, showed even greater increases, with a 17-minute increase in creation time and a 26-minute increase in commenting time (however, founders are a small group and are susceptible to statistical variability).

In other words, despite AI replacing some coding and issue creation tasks, the total time spent by the entire development team has actually increased, rather than decreased—a paradoxical result.

The Interpretation: "AI Doesn't Replace Work, It Adds to It"

This report offers one interpretation of this paradoxical result: "More work requires more coordination, and this coordination is increasingly shaping the context in which agents act."

In other words, while AI speeds up and automates individual coding tasks, a new kind of "coordination cost" is created: humans review the deliverables, coordinate the actions of multiple AI agents, and maintain overall consistency. This is a practically important insight: AI is not simply replacing existing work, but rather adding a new type of work on top of it.

Tracking "From Token Consumption to Results": The Next Step

In this report, Linear also outlines plans for further in-depth analysis. They state that with code and code reviews now passing through Linear, they can now observe "the entire lifecycle of work, from token consumption to actual results."

This suggests the potential for metric evaluation of AI coding agent effectiveness to evolve from superficial activity metrics like "adoption rates" and "number of pull requests" to more fundamental efficiency metrics such as "how much real value was generated relative to the computational cost invested."

What Researchers Should Note

Reports based on this type of real-world product usage data are a valuable source of information that reflects a more grounded reality across the industry, unlike the flashy success stories published by individual companies. While the adoption of AI coding agents certainly boosts development output, the finding that this doesn't necessarily translate directly into "reduced human work time" demonstrates that simply using "efficiency" as a yardstick is insufficient when evaluating the effectiveness of AI adoption.

As the scope of tasks handled by AI agents expands further, it will be interesting to see how much this new type of work—"adjustment costs"—will continue to increase, or at some point plateau. Continuous observation based on this kind of real-world data will likely become increasingly important in understanding how AI and humans can collaborate.

Source: linear.app
AIコーディング生産性データ分析ソフトウェア開発AI/ML
Advertisement300 × 250