AI Models as a Security Threat and New Tools for Agents

Today’s roundup: why Satya Nadella suggests assuming models are compromised and what the community is discussing, including tools and experiments.

🔥 Hot:

🔹 Satya Nadella called for treating all AI models as compromised: With this approach, model security becomes the top priority during development and use.

➡️ News:

🔹 Europe is confident its rules can contain dangerous AI 🔹 GSD Task Manager adds a task list for an AI agent via MCP

➡️ Useful materials:

🔹 A tracker of lawsuits against AI companies has been launched: The website brings all the cases together in one place. 🔹 An author is experimenting with turning dense models into Mixture-of-Experts: The idea is that part of the MLP runs on every token, and the model does not need to be trained from scratch.

➡️ Discussions and case studies:

🔹 AI agents held a demonstration stand-up 🔹 An enthusiast built a cluster for running local models on BC-250 boards: He explains how to run models via llama.cpp and Vulkan on four boards. 🔹 DeepSeek Harness helped a user restore a broken GNOME extension

📝 If you would like to add other news and materials to the list, write in the comments.