Microsoft's "Copilot OS" Revealed in Leaked Video — A Windows Rebuilt Around Copilot and Agentic AI
According to a report from Windows Central, a leaked video has revealed that Microsoft is developing an experimental prototype of a lightweight version of Windows called "Copilot OS." The OS is said to completely redesign the traditional desktop UI and is built around Copilot and agentic AI.
The leaked footage suggests that the design routes even basic operations such as launching apps and managing files through an AI assistant. It shows that the "AI-first OS experience" Microsoft has long hinted at is finally taking shape as a concrete prototype.
That said, this remains an exploratory project at this stage, and there are no known plans or details for a product release. While it is a significant signal about the future of Windows, we will need to wait for an official announcement regarding any actual release.
Ramp Data Suggests "AI Isn't Taking Jobs" — Companies Aggressively Adopting AI Are Hiring More
A Substack report analyzing Ramp's spending data found that companies aggressively adopting AI are actually increasing hiring. Amid persistent concerns that "AI will take human jobs," real-world corporate data shows the opposite trend.
What the data suggests is that companies that have improved operational efficiency through AI adoption may be directing the freed-up resources toward investments in new areas and headcount expansion. In other words, AI often functions not merely as a "headcount reduction tool" but as a catalyst for business growth.
That said, we should be careful about whether this trend applies to all industries and company sizes. Ramp's data may be skewed toward mid-market and larger tech-leaning companies, and the long-term impact will require continued monitoring.
Supply Chain Attacks on LLM Agents: Paper Demonstrates Exploitation Via Payload-Free Skills
A paper published on arXiv has introduced a new attack technique called "Payload-Less Skills" that exploits the LLM agent supply chain. The attack cleverly abuses the very mechanism of "skills" (features similar to tools and plugins) that agents rely on.
Whereas traditional supply chain attacks embed malicious payloads directly, this technique is characterized by its ability to indirectly manipulate an agent's behavior through seemingly harmless skills that carry no payload. This creates an attack vector that is difficult to detect with existing security scanners.
As the LLM agent ecosystem expands, this development once again highlights the importance of supply chain security. For teams running agents in production, verifying the trustworthiness of third-party skills and plugins has become an urgent priority.
"Kortex," an LLM Inference Engine Built from Scratch in Rust, Released on GitHub
An out-of-core LLM inference engine called "Kortex" has been released on GitHub. Written from scratch in Rust, it is designed to efficiently run large models that do not fit in memory.
Out-of-core inference is a technique that performs inference while sequentially loading model parameters from disk, with the advantage of enabling large models to run even in environments with insufficient GPU memory. For developers who want to run LLMs locally, it looks set to become a practical option.
The Rust implementation offers advantages in both memory safety and performance, reflecting a design philosophy distinct from existing C++-based engines (such as llama.cpp). With future community contributions, the options for local LLM inference are expected to expand even further.
An Australian Influencer and the Reality of Manipulation Through AI-Generated Content
Australia's ABC News has reported on how an influencer known as "Lily Jay" was spreading misleading content using AI-generated videos. The story has also drawn attention on Hacker News, gathering 7 points.
This case is a concrete example of how easily AI-generated content can spread misinformation on social media. The combination of an influencer's reach and AI's content generation capabilities makes possible disinformation campaigns that are more sophisticated and harder to detect than ever before.
As debates over platform responses and mandatory labeling of AI-generated content progress in various countries, cases like this will serve as important reference points for considering the direction of regulation.