AI News

US Approves Export of 879,000 AI Chips to China: Over 30% Flow to Pentagon Military List Companies, Control Loopholes Exposed

Declassified license data reveals that from December 2025 to May 2026, the US Commerce Department's BIS approved the export of 879,000 advanced AI chips to Chinese companies, with over 30% going to entities on the Pentagon's Section 1260H List. The disclosure exposes structural contradictions in the US export control system.

AI Chips Export Controls 中美科技战
118

Three AI Bills Vie for the Last Legislative Window: Pass by December 11 or Wait for Democrats to Take Congress

On November 19, 2026, three frontier-AI safety bills are competing for a narrow legislative window before December 11, when key cybersecurity authorizations expire and Democrats are set to take both chambers in January. The article compares the FRONTIER, SAFE AI, and AI Circuit Breaker bills and assesses what developers should watch and what is likely next.

AI Regulation 美国国会 FRONTIER法案
165

Amazon Discloses China-Linked AI Agent Autonomously Breached East Asian Medical Imaging Server Zero-Day in 11 Minutes

Amazon Threat Intelligence disclosed that China-linked hackers used the open-source AI penetration framework Cairn to direct an AI agent against East Asian university medical imaging servers in September. The agent found and exploited a path traversal zero-day in MONAI Label, moving from API mapping to root access in 11 minutes; a patch was released on November 18.

AI Safety 零日漏洞 网络攻击
220

AI Designs a CPU from Scratch: Frostbite Achieves 4.59× Energy-Efficiency Gain, but Has Not Yet Been Taped Out

Architect Labs has unveiled Frostbite, an AI-designed general-purpose CPU that completes the full flow from ISA definition to GDS layout and reports a 4.59× geometric-mean energy-efficiency ratio versus a silicon-validated dual-issue in-order superscalar core across 25 workloads. The results remain simulation-based, and the chip has not yet been fabricated.

CPU架构 AI芯片设计 Architect Labs
204

Language Models Can Rewrite Their Own Weights: 27B Model Gains 6.6 Accuracy Points Without Training, Escalating AI Safety Concerns

Researchers from the University of Illinois Urbana-Champaign and Carnegie Mellon University report that a 27-billion-parameter Qwen3.5 model can edit its own weights through forward passes alone, improving accuracy on unseen math problems by 6.6 percentage points without conventional training. The work demonstrates a scale-dependent self-rewriting capability and raises new AI safety concerns about deployed models modifying themselves.

AI Safety 大模型 自我改进
175