Kimi K3 Falls Far Behind US Frontier Models in Cyberattack Tests — Distillation to Blame?

The UK AI Safety Institute (AISI) and the US Center for AI Standards and Innovation (CAISI) have jointly published the results of testing Moonshot AI's large language model Kimi K3 on offensive cyber tasks. On ExploitBench, a benchmark designed to reproduce real-world cyberattacks, Kimi K3 managed a score of just 32%, a wide gap compared with the 76% recorded by US frontier models.

In addition, the safeguards built into Kimi K3 reportedly failed to block exploit development and simulated attacks. While Kimi K3 posts high scores on general benchmarks, it shows a striking weakness in the cyber domain, and the research team noted that this gap is consistent with claims that it stems from distillation from Anthropic's models. In other words, the data backs up the idea that distillation can boost general capabilities while leaving specialized-domain skills and safeguard implementation underdeveloped.

These results add fresh fuel to the debate over the alleged distillation behind Kimi K3.

Grok Releases an Excel Add-in — Entering the Spreadsheet Market

xAI's conversational AI Grok has released an add-in for Microsoft Excel, letting users tap Grok's capabilities directly inside Excel.

Microsoft delivers AI features across its Office suite through its own Copilot, and Grok's move into Excel is drawing attention as an effort to widen the range of available AI assistants. AI integration in spreadsheet software is highly practical for data analysis and automation, making it a fiercely contested area among major players.

Note that as of this writing, details such as the add-in's availability and pricing remain limited to what can be confirmed from restricted-access information.

Hetzner Is Developing an LLM Inference Service

German hosting provider Hetzner is reportedly developing an LLM (large language model) inference service. Hetzner has long been popular with the developer community for its affordable yet powerful dedicated servers, and its entry into the LLM inference market looks like a play that leverages the company's strength in driving down GPU infrastructure costs.

The LLM inference infrastructure market is currently expensive, centered on NVIDIA GPUs, and the arrival of low-cost providers like Hetzner could lower the barrier to entry for small businesses and individual developers. Detailed specs and launch timing will have to wait for future announcements, but the topic has already drawn 12 points of interest on Hacker News.

VideoChat3, a Fully Open-Source Video Understanding MLLM, Has Arrived

A video-focused multimodal large language model (MLLM) called VideoChat3 has been released. Its defining trait is balancing versatility and computational efficiency despite a relatively compact size of 4B parameters.

On the technical side, it introduces I3D-ViT (Inflated 3D Vision Transformer), which efficiently processes spatial and temporal representations, along with Adaptive Frame Resolution, which cuts the cost of processing streaming video. The team also built its own training data generation pipeline and constructed three datasets covering three scenarios — general-purpose, long-video, and streaming: VideoChat3-Academic2M, VideoChat3-LV116K, and VideoChat3-OL617K.

Whereas existing open-source models have tended to skew toward specific domains, VideoChat3 releases everything — training code, strategies, and datasets alike — making it a model that can genuinely contribute to community-driven reproducibility and progress.

AI-Generated Supermarket POP Signs Spark Debate — A New Side of Generative AI Adoption in Japan

Flashy product point-of-purchase (POP) signs, apparently created with generative AI, displayed in the produce section of a Japanese supermarket have become a hot topic on X (formerly Twitter). While some commenters find the designs — featuring vampires and Sengoku-era warlords — hard to read, Dr. Yasukazu Kinuta of Kinuta Dental Clinic, famous for its signboard advertising, gave them his stamp of approval, calling them 'this works.'

The episode highlights how AI-generated content is starting to be used routinely on the retail floor as generative AI design tools spread. It offers a snapshot of a transitional moment: the benefit of instantly producing professional-quality graphics coexisting with practical drawbacks such as reduced readability from over-the-top designs.

---