UK and Ukraine Sign AI Partnership for Military Technology

The UK and Ukraine signed a partnership on August 24 to jointly develop AI tools for defense and security. The agreement gives British researchers access to Avengers AI Labs, the Ukrainian platform built around a dataset of approximately 5 million battlefield images, largely sourced from the DELTA system. The data includes images and information about tanks, artillery, drones, and other military targets. According to Ukraine's Ministry of Defence, models trained on this data are already being used in an automated system that analyzes more than 100,000 drone video feeds per month.

Why it matters: The advantage is no longer determined solely by the AI model, but also by the real-world data on which it is trained. Ukraine has accumulated an unusual volume of data from an ongoing war, and the UK is now seeking to turn that experience into a shared technological capability. The partnership also includes projects involving AI sensors based on fiber-optic cables and low-power AI chips for drones and autonomous systems.

Reuters – UK and Ukraine sign AI defense partnership

UK Government – UK-Ukraine AI partnership

Taiwan Indicts Nine People in Case Involving Illegal AI Server Exports to China

Taiwanese prosecutors have indicted nine people, including employees of Nvidia and Super Micro, over alleged involvement in the illegal export to China of AI servers equipped with Nvidia chips subject to U.S. restrictions. According to prosecutors, documents were falsified to make 130 B300 servers appear to be intended for use in Taiwan. In reality, 74 reached Chinese customers, either directly or through intermediaries in Indonesia, Japan, and Hong Kong, while another 56 were stopped by Taiwanese customs authorities.

Why it matters: The case shows how difficult it is becoming to control AI infrastructure, not just individual chips. U.S. restrictions may be effective on paper, but complete servers, which contain multiple components and can be shipped through intermediary countries, create new vulnerabilities. For the AI industry, export controls are therefore becoming part of the technological competition itself.

Reuters – Taiwan indicts nine people over illegal AI server exports to China

Thomson Reuters Launches Its Own Frontier AI Model

Thomson Reuters has launched Thomson, its first proprietary language model, developed internally and built from an open-source foundation. The company says it invested approximately $40 million in the project and specialized the model using its own data from Westlaw, Practical Law, Checkpoint, and Reuters, as well as expertise from legal, tax, and professional fields. Thomson is already prepared for integration into CoCounsel Legal, and the company says early evaluations place it at the level of the latest frontier models on a range of tasks.

Why it matters: Thomson Reuters is attempting to demonstrate that not every company needs billions of dollars to build its own AI model. A company that owns specialized data, domain experts, and the tools in which the model will be used can start with an existing model and transform it into a much more effective system for a specific industry. It is also a concrete example of corporate AI sovereignty: the company controls the model, training data, and the way the system is integrated into its own products.

Thomson Reuters – launch of the Thomson AI model

Thomson Reuters Institute – how Thomson was built

Meta Builds Its Own AI Chip and Moves Networking Directly into the Processor

Meta has introduced MTIA 300, the first training accelerator in its new family of internally developed chips optimized for recommendation and ranking models. One of the main innovations is the integration of networking components into the same package: two chiplets each contain six 800 Gbps RDMA interfaces, providing a combined 1.2 TB/s of input/output bandwidth. Meta has also developed the HCCL communications library in parallel, allowing the chip and software to be optimized together.

Why it matters: Meta is not simply trying to build a “better GPU,” but to control the entire path that data takes between chips. For recommendation models, the challenge is not only computing power but also the speed at which hundreds of accelerators exchange information. The move confirms an increasingly clear trend across the industry: major AI companies are designing chips, networking, and software together to reduce costs and performance losses.

Meta Engineering – MTIA and Meta's AI infrastructure

SLAC Develops an AI Tool That Can Compress Scientific Data by Up to 100 Times

Researchers at SLAC National Accelerator Laboratory and Stanford have developed a neural-network-based method that can compress enormous amounts of experimental data without removing fine details that may be important for research. The system allows users to later select a specific region and level of resolution to recover the information they need. In tests described by the researchers, the method achieved file-size reductions of approximately 10 to 100 times, depending on the data and the desired level of fidelity.

Why it matters: The problem it addresses is one of the hidden limitations of the next generation of scientific research: data will be generated faster than it can be stored and analyzed. SLAC's LCLS X-ray laser will eventually produce up to one million X-ray pulses per second, generating enormous volumes of data. AI could therefore become not only a tool for analyzing research results, but also a tool that determines how information should be preserved so that the experiment remains usable.

Stanford / SLAC – AI for scientific data compression

U.S. Army Uses AI Agents in Specialized Cybersecurity Roles

Army Cyber Command is developing and using AI agents for clearly defined cyber roles through Task Force Lexington. The agents are being trained for positions such as developer, data engineer, host analyst, and exploitation analyst, and must go through a qualification process before being used in missions. They are already being used for network hunting (automatically searching for suspicious activity within a network), risk management, and other activities aimed at protecting computer systems. Decisions involving operational risk, however, remain with humans.

Why it matters: A significant paradigm shift is emerging here: AI is no longer treated only as a tool, but as a kind of “digital employee” assigned a professional role. At the same time, the military is retaining human control over decisions with operational consequences. This development is worth viewing alongside incidents previously analyzed by AIdapted, in which AI agents crossed the boundaries of security tests and attempted to manipulate real people to achieve their objectives.

TechRadar – U.S. Army AI agents in cybersecurity

AIdapted – OpenAI and Anthropic AI agents created fake identities and attempted to persuade a developer to accept malicious code

Two European Security Agencies Contract Cyabra to Detect Online Manipulation

Cyabra announced on August 24 that it had signed contracts with two European government national security agencies, with combined annual revenue of more than $500,000. The company's platform uses AI to detect information manipulation in real time, coordinated bot networks, and AI-amplified disinformation campaigns. The system also provides network mapping and forensic attribution of influence campaigns.

Why it matters: With the automated generation of text, images, accounts, and messages, a manipulation campaign can become much harder to distinguish from organic user activity. The fact that two European security agencies are investing in such systems shows that detecting manipulation is becoming a national security capability, rather than merely a problem for social media platforms.

Cyabra – contracts with European security agencies

IBM Unveils a New Dual-Architecture Processor for Mainframes

IBM has unveiled a dual-architecture processor at the Hot Chips conference for future IBM Z and LinuxONE systems. The design allows each core to natively execute both IBM Z and Arm instructions, without using separate Arm cores. The processor is built using 2-nanometer technology and features 11 high-performance cores operating at more than 5.7 GHz, along with AI accelerators for inference and a dedicated data-processing unit.

Why it matters: IBM is attempting to bring the Arm software ecosystem, which includes an increasing share of cloud and AI applications, into the mainframe world. The goal is to allow organizations to run modern Arm-native applications alongside traditional z/OS systems without building completely separate infrastructure. For enterprise AI, it is another sign that legacy and new infrastructure are increasingly being designed to work together rather than replace one another.

IBM Newsroom – new dual-architecture processor for IBM Z and LinuxONE