#ai-safety

AI agents are increasingly acting on real infrastructure without adequate authorization checks, from a gym-booking exploit to an unresolved Hugging Face attack timeline. Economists and researchers are meanwhile debating whether to slow AI self-improvement through regulation, with dueling arguments this week over data center bans and safety-rule timing. Elsewhere, GitHub retired its unified LLM API, and the FDA rejected a manufacturing-flagged cancer drug.

The FDA's leadership vacuum deepened this week as the drug center head followed Commissioner Makary out the door, leaving the agency with no confirmed leadership at its two most critical posts. Open-weight models kept multiplying with Gemma 4, DeepSeek V4, and Kimi K2.6 all dropping in the same month, while Sebastian Raschka's analysis of new KV-sharing and compressed attention techniques explains why long-context costs are falling fast. A rare Bundibugyo Ebola outbreak in Congo's Ituri province prompted a WHO public health emergency declaration, and the WHO's move landed as hantavirus conspiracy theories were already cycling through social media. On the economy side, Trump's China visit produced mixed signals on decoupling, while Noah Smith argued that Europe's comfortable stagnation is no longer viable.

Sebastian Raschka published a comprehensive 2025 LLM review covering DeepSeek R1, RLVR, inference-time scaling, and 2026 predictions, while Simon Willison released his annual year-in-LLMs retrospective. On the pharma front, NIH Director Jay Bhattacharya confirmed that DEI-related grants restored under a court order will not be renewed in 2026, even as NIH began reviewing thousands of delayed proposals and funded 135 on the first day. Novo Nordisk priced its newly approved Wegovy pill below Eli Lilly's competing obesity drug in a direct-to-consumer push. On the economy side, researchers across 16 countries found that the post-1980 labor share decline is driven by the composition of capital, not its quantity, and building more structures may reverse it.