🏛️ White House pushes OpenAI to delay GPT-5.6 release 🏛️
The White House demands a staggered rollout for OpenAI's most capable model yet.

The White House has formally requested OpenAI to delay the public deployment of GPT-5.6, citing national security and structural safety concerns. Under the proposed plan, the model will first be previewed to a small group of government-approved partners on a customer-by-customer basis, while officials push for an extended red-teaming window to audit the system's advanced cyber-capability limits and automated social manipulation vulnerabilities — marking a direct escalation in government regulation of frontier AI. CEO Sam Altman told employees this path is the most likely route to a full public launch, expected "a couple of weeks later." OpenAI has made clear that this kind of gated rollout is not its preferred long-term model.

🤖 Meta ships Autodata for better training data 🤖
Meta's new tool accelerates dataset creation with AI-powered data agents.

Meta has released Autodata, a new tool that trains AI agents to function as automated data scientists, generating higher-quality training and evaluation datasets at speed. The system's Agentic Self-Instruct implementation demonstrated measurable improvements across coding, legal reasoning, and mathematical reasoning benchmarks, shifting competitive pressure from parameter scale to data quality. The approach could significantly reduce the manual effort required to curate specialized training datasets. The research was published on arXiv, making the full methodology available for scrutiny by the broader research community. It underscores how frontier labs are increasingly leveraging agentic AI to improve the AI development pipeline itself.

📊 Microsoft adds AI Skills to Excel Copilot 📊
Microsoft deepens Copilot integration in Excel with reusable AI-powered workflows.

Microsoft has introduced AI Skills for Copilot in Excel, a new capability enabling reusable AI-powered workflows directly within spreadsheets. The release includes pre-built skills targeting financial modeling, forecasting, and variance analysis — core enterprise use cases that move Copilot beyond simple Q&A into structured, repeatable operations. By offering domain-specific pre-built skills, Microsoft lowers the barrier for business users to adopt AI in data-heavy workflows, a strong adoption signal for enterprise AI. The feature builds on Microsoft's broader Copilot strategy across Microsoft 365, with the announcement on the official 365 blog signaling general availability.

INTERESTING TO KNOW

🚨 Anthropic flags Alibaba for 28.8M Claude data theft 🚨

In a disclosure that could reshape API access policies across the industry, Anthropic has accused Alibaba of conducting the largest known distillation attack against Claude. Nearly 25,000 fraudulent accounts extracted 28.8 million exchanges over 45 days, specifically targeting Claude's agentic reasoning and coding capabilities. Anthropic disclosed the operation in a letter to the Senate Banking Committee, calling for stronger chip export controls and sanctions against Chinese labs. This follows earlier attacks by DeepSeek, Moonshot, and MiniMax totaling over 16 million exchanges.

💻 DeepReinforce drops Ornith-1.0 coding models 💻

DeepReinforce has released Ornith-1.0, a family of open-source coding models that write their own reinforcement learning scaffolds — a notable speed gain in autonomous model development. Built on Gemma 4 and Qwen 3.5 foundations, the models reportedly rival Anthropic's Opus 4.7 on coding benchmarks, closing the competitive gap between open-source and proprietary frontier models. All weights and a detailed technical report are available on Hugging Face.