Anthropic Wins Court Battle Against Pentagon After Being Declared a “Supply-Chain Risk”

A federal judge in California ruled that the U.S. administration acted unlawfully when it designated Anthropic as a “supply-chain risk,” a classification that led the U.S. government to restrict cooperation with the company. According to Reuters, Judge Rita Lin found that the measure amounted to unlawful retaliation and violated the company’s constitutional rights. The dispute began after Anthropic refused to remove certain restrictions on the use of Claude models for fully autonomous weapons and mass surveillance.

Why it matters: The ruling could set an important precedent for the relationship between AI companies and governments, particularly when model providers reject certain military or security applications. The case is even more relevant after Anthropic previously warned about the risks of using AI without adequate safeguards; AIdapted has already followed this dispute in its article about the limited restoration of U.S. access to Claude Mythos 5.

Russian Hackers Used Cursor AI Agent to Attack at Least Seven Companies

A Russian-speaking hacking group associated with ransomware operations used the AI agent in Cursor to accelerate attacks against several companies. According to Reuters, researchers at Gambit Security identified more than 20 conversation sessions in which the attackers convinced the agent that they were conducting legitimate security tests. AI was used for activities including network scanning, privilege identification, VPN access configuration, and attempted exploitation.

Why it matters: The incident highlights a different problem from conventional AI-generated attacks: criminals do not necessarily need to build their own offensive AI system. They can use commercial tools created for legitimate developers. This is particularly relevant to the discussion about AI agents that AIdapted has already covered in its article about agents that created fake identities and attempted to introduce malicious software. As agents receive greater access to real systems, the distinction between an “assistant” and an “operator” becomes increasingly blurred.

Cisco Gives Every One of Its 90,000 Employees a Personal AI Agent

Cisco is expanding MyAgent, its own AI-agent system, to its entire workforce of approximately 90,000 employees. The company describes the system as a form of “ambient intelligence” that can retain context, understand user intent, and move beyond conversation toward actually executing tasks. Cisco says the rollout is part of its strategy to integrate AI agents directly into internal workflows.

Why it matters: Cisco’s experiment is one of the largest corporate tests yet of the idea that every employee can have their own “digital colleague.” If the model works, enterprise AI could move beyond licenses for chatbots and copilots toward agents that actually execute processes, use internal data, and interact with multiple company systems.

China Opens Government Intellectual Property Data for Commercial AI Use

China is preparing to expand access to government-held intellectual property data, including for commercial applications and AI development. The information was reported by MLex, as part of a broader strategy through which Beijing is attempting to turn public-sector data into a resource for technological development. At the same time, China’s intellectual-property authority is developing rules for protecting and commercializing data rights, including in areas such as AI and “embodied intelligence.”

Why it matters: For AI models, access to large, high-quality datasets can become a competitive advantage comparable to access to chips and data centers. If government data can legally be used at commercial scale, Chinese companies could gain an additional source for training and specializing their models.

Tencent Launches Hy4 Preview, a New Open-Source AI Model with 770 Billion Parameters

Tencent launched Hy4 Preview on August 28, an open-source model with 770 billion parameters, approximately 49 billion of which are activated for each token. The model has a context window of more than one million tokens and is designed for coding, research, data analysis, and other complex tasks. Reuters reports that the model is part of the increasingly aggressive push by Chinese companies into the open-model segment.

Why it matters: The model’s total size is impressive, but more significant is the fact that Tencent activates only part of its parameters for each operation, potentially reducing inference costs. Hy4 also illustrates how quickly Chinese open-source models are approaching the capabilities of Western systems, in a competition where price, model accessibility, and infrastructure efficiency are becoming increasingly important.

DeepSeek Heads Toward a Valuation of Around $74 Billion

Chinese startup DeepSeek is approaching a new funding round that could value the company at approximately $74 billion. South China Morning Post reports that the company is discussing a round of around 50 billion yuan, while sources cited by the Wall Street Journal point to a valuation of roughly $74 billion and plans to expand research and computing infrastructure. The company is also considering a potential listing on Shanghai’s STAR Market.

Why it matters: The valuation shows that DeepSeek’s success is no longer viewed solely through the performance of its low-cost models, but also as an AI business capable of attracting capital on a multibillion-dollar scale. It also represents another stage in China’s push to develop its own infrastructure and models, a theme AIdapted has previously examined in its coverage of Chinese models and the distillation of Western AI systems.

Chinese AI Agents Are Beginning to Be Used in Cyberespionage Operations

The use of AI agents in cyberespionage operations is becoming an increasingly concrete concern. Recent analyses show that China-linked groups have experimented with AI models to automate stages such as infrastructure reconnaissance, vulnerability identification, exploit development, and data extraction. Carnegie Endowment notes that, in an operation attributed to a Chinese group, Claude was used to automate a large part of the intrusion process, while Google and OpenAI have observed their own systems being used by state actors for offensive activities.

Why it matters: the shift is from AI being used as a targeted tool to AI being used as an operator across multiple consecutive stages of an attack. AIdapted has already documented the autonomous and deceptive behavior of AI agents in real-world security tests, while cyberespionage cases show why these tests are no longer merely academic exercises.

Germany Prepares a Data Center of Up to €5.6 Billion for Cloud and AI

Schwarz Group and the state of Mecklenburg-Vorpommern have announced a partnership that could lead to an investment of up to €5.6 billion by 2033 to build a data center dedicated to cloud and AI. According to Schwarz Group, the project is designed to strengthen Germany's digital infrastructure and enable the development of cloud and AI services under German law. The plans include the use of renewable energy and the integration of heat generated by the infrastructure into a local heating network.

Why it matters: the investment shows that Europe is beginning to treat AI data centers as strategic infrastructure rather than simply commercial properties. The trend is also relevant to Romania: AIdapted has already highlighted in its analysis of RO AI Factory and in its article on new European locations for AI infrastructure that energy availability and grid connectivity are becoming decisive factors when choosing locations.

Romania Prepares Two Data Center Projects with a Combined 380 MW Capacity in Ilfov

WDP Romania and Lamas Invest have begun procedures to develop two data centers in Ștefăneștii de Jos, Ilfov County. The projects are currently in the permitting and grid-connection stage and have received approvals for a combined power consumption of 380 MW, with 190 MW allocated to each center, according to information published by Profit.ro. A 220 kV transformer substation is also planned to supply the facilities, while the necessary energy infrastructure is expected to become operational by the end of 2030. Profit.ro – the Ilfov data center projects

Why it matters: the 380 MW represents a very large power requirement for data center infrastructure and shows how important energy availability has become for the development of AI and cloud services. However, these are not two data centers that have already been built, nor is this a completed investment. They are projects under development for which the necessary power-consumption approvals have been obtained, while permitting and grid connection are still underway. If completed, the projects could make a significant contribution to the infrastructure needed for cloud services and AI applications in Romania.

Europe Invests in Specialized AI: From Healthcare Agents to Cognitive Robots

Dutch startup Clementine, based in Maastricht, has raised €1.7 million in a Seed round to develop its AI platform for audiology clinics. The system includes an agent called “Matthew,” which can communicate continuously with potential patients, answer questions, guide them toward appropriate services, and turn conversations into appointments. At the same time, German robotics startup NEURA Robotics announced a Series C round of up to €1.2 billion, backed by investors including Tether, Qualcomm, Amazon, Nvidia, Bosch, and the European Investment Bank. The company is using the funding to develop its “Neuraverse” platform and expand production of cognitive robots, and strengthened its position in August through the acquisition of Adlatus Robotics.

Why it matters: the two investments illustrate two different but complementary directions for the next stage of AI. Clementine shows how AI agents are being adapted to highly specific processes with immediate commercial impact, while NEURA Robotics is pursuing Physical AI — systems capable of perceiving and acting in the physical world. The distinction is important: AI is no longer limited to models that generate text or images, but is increasingly being integrated into interactions with patients, robots, and real-world processes. The scale of NEURA's funding also shows that European and international investors see robotics as one of the next major markets for AI.

Spain Opens Consultation on Stricter Rules for the Use of Generative AI in the Arts Industry

Spain’s Ministry of Labour has launched a public consultation for a future organic law governing the use of generative AI in employment relationships within the arts industry. According to DLA Piper’s analysis of the Spanish initiative, the proposal aims to establish rules governing the use of artists’ images, voices, and other recognizable elements in AI-generated content. In principle, the proposal includes requirements concerning consent, contractual limitations, and, in certain circumstances, separate compensation.

Why it matters: Spain is seeking to turn the issue of AI-generated replicas of artists into an explicit matter of employment law and the protection of fundamental rights. If this approach spreads, companies using AI to generate virtual actors, voices, or images of real people may need not only technical licenses, but also clear contractual agreements.

Sources