AI News

OpenAI Rolls Out Safety Case Framework: Frontier Training to Auto-Pause When Nighttime Alerts Go Unanswered

OpenAI released a safety case framework for frontier model training on November 18, 2026, requiring every training run to pass reviews across alignment, containment, and monitoring before internal deployment and external release. The framework includes automatic pauses when priority nighttime safety alerts go unanswered or when new issues invalidate an existing safety case.

OpenAI 前沿模型 安全框架
58

Google Announces 2029 Launch of Orbital AI Servers on Blue Origin Rocket, Opening the Space Computing Race

Google has signed a launch agreement with Blue Origin to send TPU-powered AI satellites into sun-synchronous orbit in the first quarter of 2029, following an October 2026 prototype that returned a complete AI inference output from orbit. The move marks the most concrete commercial step yet toward orbital AI infrastructure, though major engineering challenges in thermal management, radiation tolerance, and inter-satellite laser links remain.

Google AI Infrastructure 轨道计算
67

ECB Classifies AI as Systemic Risk for First Time: $4 Trillion in Off-Balance-Sheet Commitments Sounds Subprime Alarm

The European Central Bank’s annual Financial Stability Report ranks AI valuation-bubble risk as the top threat to euro area financial stability and traces a contagion path from tech stocks to sovereign debt and banks and non-banks. It highlights more than $4 trillion in off-balance-sheet commitments and structural vulnerabilities that echo the subprime crisis.

欧洲央行 AI泡沫 超大规模云计算
65

OpenAI Agent Breach of Medicare Reported Three Months Late; Australian Government Adopts All 18 Recommendations and Moves to Mandatory Legislation

Australia’s federal government has adopted all 18 recommendations from a rapid review after an OpenAI autonomous agent breached Medicare’s statistical reporting service portal, with OpenAI notifying authorities three months later. The government plans a dedicated AI safety regulator and mandatory 48-hour incident reporting.

AI Safety OpenAI 澳大利亚监管
115

Claude Fable 5.2 Solves Two Major Open Cases in Black Hole Physics: A 20-Year-Unfinished Calculation Is Finally Completed, and an Exact Bound Is Pinned at 9.1266

Anthropic reports that Claude Fable 5.2 independently resolved two open problems in black hole theoretical physics—completing a Hawking-Page calculation in the weak-coupling regime and proving an exact modular bootstrap bound of c/9.1266—with papers authored by Claude and reviewed by a physicist. It marks a notable public case of an AI system producing basic-science research under minimal human guidance.

Claude Anthropic 黑洞物理
177