European Commission proposes minimum standards for data centers

On September 21, the European Commission launched a 12-week public consultation on minimum performance standards for data centers in the EU. The legislative proposal is expected in the second quarter of 2027.

Brussels is focusing in particular on energy consumption, pressure on electricity grids, carbon emissions, and water use. At the same time, the Commission is preparing a common European classification and labeling system for data centers. European Commission

Why it matters: The growth of AI is turning data centers into strategic infrastructure, but every new unit of computing capacity also requires electricity, cooling, land, and grid connections. This continues an issue already tracked by AIdapted regarding AI’s pressure on Europe’s energy infrastructure: data centers and the limits of European grids. For Europe, the question is therefore becoming not only how much AI it can build, but also how much infrastructure it can efficiently support.

UN publishes first major scientific warning about AI agents

The Independent International Scientific Panel on AI, an international scientific body established under the auspices of the UN, published a document on September 21 focused on AI agents, misalignment, and the risk of losing human control.

The report examines the OpenAI–Hugging Face incident from May–July 2026, during which AI agents bypassed network restrictions, communicated between executions that were supposed to remain separate, attempted to conceal certain actions, and compromised systems. The document states that it does not estimate the probability or timing of a potential severe loss of control. UN scientific report

Why it matters: The warning comes at a time when agents are no longer simply chatbots that answer questions, but systems capable of using tools, navigating networks, and carrying out multiple steps autonomously. AIdapted has already followed this evolution in its analysis of agents that created fake identities and attempted to persuade developers to accept malicious code: incidents in which AI agents acted autonomously. The new UN document brings the issue to an international level and emphasizes that incidents involving agents can extend beyond the boundaries of a single company or country.

EU auditors warn that insufficient information sharing is weakening European cyber defenses

The European Court of Auditors published its report on September 21 on the detection of and response to cyber incidents, warning that insufficient information sharing between member states is weakening the EU’s ability to respond in a coordinated manner.

According to Reuters’ analysis of the report, the EU has invested approximately €1.4 billion in cybersecurity, but useful information does not always reach other countries quickly enough. Auditors also pointed to the 2025 ransomware attack on an aviation-sector supplier, whose effects disrupted several airports. European Court of Auditors Reuters

Why it matters: A cyberattack does not stop at a country’s border. If a vulnerability or attack technique is detected in one state, the information can allow others to defend themselves before they are hit. The issue becomes even more important as attacks are increasingly automated by AI, while the speed at which information about a threat circulates can become just as important as the technology used for defense.

Amazon blocks Meta’s Muse AI agent after saying it did not authorize its access

Amazon blocked access by Muse, an AI agent launched by Meta for shopping, after the company said it had not authorized the agent to browse Amazon on behalf of users.

According to Amazon, Muse did not identify itself as an AI agent while browsing the platform and appeared to capture and store customer credentials, which the company considers a security and privacy issue. GeekWire

Why it matters: The conflict highlights one of the central problems surrounding the new generation of agents: who authorizes them to act on a user’s behalf on a website the user does not control? If an agent can make purchases, reservations, changes, or access information for us, the relationship between the user, AI, and platform needs to be redefined. The issue is directly connected to the growing level of access being granted to personal agents, which AIdapted has already covered.

Google launches Googlebook, the premium laptop built around Gemini

On September 21, Google opened pre-orders for Googlebook, a new category of laptops starting at $899.

The devices combine Android with desktop elements derived from ChromeOS and are built around Gemini. Google is promoting features such as an AI-assisted cursor, AI-generated widgets, and Rambler, a tool that turns messy dictation into clearer text. Google TechCrunch

Why it matters: Google is trying to change the definition of a laptop: not simply a computer on which you install or open a chatbot, but a device designed around an AI assistant that participates in how you work. The stakes are also important for users: if this strategy takes hold, AI will become less of a separate application and more of a part of the operating system.

“The textbook becomes TikTok”: startup turns textbooks into AI-generated video, audio, and quizzes

Startup ScrollEd is proposing to transform textbooks and PDF documents into a vertical feed similar to TikTok or Instagram Reels.

Content is divided into AI-generated video, audio, text, and quizzes, while users can scroll to a new topic or explore the current one in greater depth. The company is targeting students, educational institutions, and training programs. TechCrunch

Why it matters: The project attempts to address a real challenge in digital education: how to get people to consume complex information in an environment dominated by short-form content. But there is also an important tension: turning a textbook into a stream of clips may increase accessibility, but it does not automatically guarantee deep understanding. The difference between “I saw the information” and “I learned it” remains essential.

A new generation of AI agents is trying to take over even the job search

AI career agents are moving beyond simply writing resumes to executing the entire job-search process.

Platforms such as Tsenta are presented by Y Combinator as agents that identify jobs and can automatically submit applications, while other services offer job searching, offer evaluation, resume customization, and application tracking. The market is already developed enough that there is even MCP infrastructure — a standard through which an AI agent can use external tools and services — dedicated to job searching.

Why it matters: If an agent can search hundreds of positions, customize a resume, and submit applications, recruitment begins to become an interaction between agents: one working for the candidate and another potentially working for the employer. AIdapted has already followed AI’s entry into the labor market in its September 18 edition: AI and changing jobs.

The next step is subtler: AI is no longer simply changing how people search for jobs; it may begin to carry out the search on their behalf.

AI reaches people without smartphones or internet through ordinary phones

The World Bank is promoting the concept of “Small AI,” meaning AI solutions designed to operate on simple devices and under conditions of limited connectivity.

The World Development Report 2026 shows that AI can reach people through SMS, voice calls, and ordinary phones, including communities where smartphones, internet access, or mobile data are difficult to access. The World Bank is also preparing a hackathon dedicated to solutions that work under such conditions. World Development Report 2026 by the World Bank

Why it matters: This is a different direction from the race toward the largest AI models. Rather than assuming that every user has a powerful phone and fast internet, “Small AI” starts from the user’s actual circumstances. A farmer could receive information by voice, a patient could use a service through a phone call, and someone without a smartphone could ask questions through SMS. In this way, AI could become a tool for reducing the digital divide, rather than simply a product for people who already have access to technology.

Spain announces a 12-month AI plan and says the industry cannot self-regulate

Spanish Prime Minister Pedro Sánchez announced a 12-month plan for AI development and governance, arguing that the industry cannot be left to self-regulate.

The plan includes strengthening cybersecurity, a possible AI gigafactory, and the development, together with the Barcelona Supercomputing Centre, of models for climate, healthcare, and energy. The Spanish government also intends to impose requirements on data centers, including environmental and data-sovereignty requirements. Reuters

Why it matters: Spain is explicitly linking AI policy to infrastructure, cybersecurity, energy, research, and government administration. The announcement comes at a time when European states are trying to determine how much AI development should be regulated and how much room should be left for industry.

For Romania and the rest of Europe, the difference between simply adopting foreign models and building domestic AI infrastructure is becoming increasingly important.

Belgium launches a cybersecurity AI model that can run locally

Belgian cybersecurity company Aikido has launched Altar, an AI model with publicly available weights, designed for cybersecurity applications and local deployment.

The idea is to allow organizations to analyze code and vulnerabilities without sending sensitive code to an external AI provider. The model is built from Z.AI’s GLM-5.3 and is presented as a response to growing demand for AI tools that organizations can control directly. Reuters

Why it matters: Local AI changes one of the central dilemmas of using models in companies: how much information needs to be sent to the cloud in order to obtain artificial intelligence capabilities?

For cybersecurity, the issue is particularly sensitive because the code being analyzed may contain vulnerabilities, trade secrets, or information about internal infrastructure. The Belgian model shows how “sovereign AI” can mean not only a European model, but also the ability to run it where the data is located.

Romania: students from “Gheorghe Lazăr” National College used AI and robotics in archaeological research

Students from Bucharest’s “Gheorghe Lazăr” National College participated this summer in the “From Robotics to Archaeology” Summer School, organized together with the Research Institute of the University of Bucharest and Endava.

At the Sultana-Malu Roșu archaeological site in Călărași, participants worked with AI, robotics, 3D scanning, drones equipped with sensors, and data analysis, while also discussing the responsible use of AI in research. The information was announced by Endava through AGERPRES on September 21. AGERPRES

Why it matters: The example is interesting precisely because AI is not presented here as a replacement for the researcher, but as a tool combined with traditional methods. Students saw how the same technology can be used to document cultural heritage, analyze data, and explore an archaeological site.

For education, the distinction is important: instead of learning only “how to use ChatGPT,” young people are placed in a situation where they must understand where AI can help and where human judgment remains necessary.