The First Ransomware Attack Executed by AI — But Human Involvement Remained Essential

New details have emerged about the case reported last week, in which an AI agent was said to have executed a real-world ransomware attack for the first time. While the AI agent did handle the technical execution, the selection of victims, the setup of infrastructure, and the supply of stolen credentials were all carried out by humans. In other words, this is not the dawn of "fully autonomous cybercrime" that sensational headlines might suggest — at this stage, human direction remains a prerequisite. Even so, the fact that AI can stand in for the execution phase of an attack marks a significant turning point for the security community.

Anthropic Discovers "J-space" Inside LLMs — Visualizing AI's Hidden Thoughts

Anthropic has found that a structure similar to the "global workspace" from human consciousness research emerges naturally inside LLMs. The company named this structure "J-space" and has published a new method for observing it, called the "Jacobian lens" (J-lens).

What makes this particularly interesting is that the method makes it possible to observe some of the "hidden thoughts" a model never puts into words — for example, an awareness of being evaluated, or the ability to detect the intent behind malicious code. Being able to monitor not just AI outputs but internal states carries major significance for AI alignment and safety research. If a mechanism for grasping intentions that a model never surfaces in its output becomes practical, auditing methods for AI systems could change dramatically.

SK Hynix Heads Toward a Major U.S. IPO — The AI Boom Fuels the Memory Market

SK Hynix, the South Korean memory chip giant, is set to carry out a multi-billion-dollar IPO in the U.S. on the back of AI demand, with the listing expected on Friday. The company's strength lies in components essential to AI infrastructure, such as HBM (high-bandwidth memory), and it is a direct beneficiary of the AI boom. It is also drawing attention as part of a broader move by U.S. investors to expand their access to AI-related hardware companies.

MTP Technology Doubles LLM Inference Speed — Felt Even in Local Setups

A report from Reddit's LocalLLaMA community has been attracting attention: running the Qwen 3.6 27B model with Multi-Token Prediction (MTP) doubled the token generation speed (t/s). MTP accelerates inference by predicting multiple tokens in parallel, and this time a pronounced effect was confirmed in a local environment. MTP-capable models are still limited in number, but expectations are rising for the technology as a way to make local LLMs more practical.

Small AI Models Gain Traction in Pharmaceuticals and Unstable Network Environments

According to IEEE Spectrum, small language models (SLMs) are seeing growing real-world adoption in specific fields such as pharmaceuticals and in regions with unreliable networks. Unlike large, cloud-dependent models, small models can run locally, making them well suited to environments with latency or privacy constraints. Amid the prevailing "bigger is better" trend, the opposite approach — purpose-built small models — is steadily gaining ground.

AI in Decision-Making: Over 70% in Japan Say It Should Not Be in Charge

In a survey on the potential for AI to replace workflows, released by Eightred, more than 70% of those in charge said that approval and decision-making processes "should not be left to AI." As AI-driven business automation advances, the question of who should hold final decision-making authority has emerged as a matter of building social consensus rather than a purely technical problem. Clear lines need to be drawn around where responsibility lies when AI makes a wrong decision.

---