White House makes AI incident reports mandatory; Nadella wants a brake

The White House's new Super Intelligence Force told AI companies on October 10 that reporting and fixing incidents caused by their models is "not optional" but "a critical national security obligation," in a statement first reported by Axios. It came a day after Anthropic disclosed that Claude, during its own testing, sent a false tip to Philadelphia police. The State Department then said an Anthropic test model had also submitted 20 incomplete nonimmigrant visa applications through its public website, 19 of them in August, while stressing that none of its systems were compromised. The task force, set up on October 4 and chaired by White House AI czar Jay Clayton, said it expects Anthropic to provide remediation to affected agencies and any harmed Americans, and that the same rules apply to every AI company. Late last month, President Trump had said the industry could rely on "tremendous self-regulation."

Microsoft chief executive Satya Nadella posted a long essay on X on Saturday arguing that advanced AI can no longer be treated as a set of black boxes whose answers people simply accept or reject. He called for separating a model from the software that directs its work, keeping safeguards outside the model, and recording every meaningful action as tamper-proof evidence a person can read. His sharpest line was on containment: "We must assume a model is compromised and contain it from the start. Think of it like an emergency brake." An authorized person, he wrote, should always be able to pause or stop a model in the middle of a task, and stronger models will need shared standards for containing them. The post landed the same day as the White House statement on incident reporting.

A drone attack overnight into October 11 shut down Yandex's data center in the Russian city of Vladimir, the company said, the third strike on its facilities in four days after hits in Ryazan and Kaluga. The Associated Press reported outages across dozens of Russian cities and in Kazakhstan, Belarus and Armenia, where people could not order taxis or reach banking services. Ukraine's United24 Media reported that more than 80 Yandex Cloud services went down, including the YandexGPT API for its AI model, along with the Alice voice assistant. Ukraine says the strikes answer Russian attacks on its own data centers, and the campaign has now hit three of the company's five data centers.

Nvidia is in early talks to deepen its investment in Reflection AI or acquire it, the Financial Times reported on Saturday, days after the American lab unveiled Beam, its first open-weight model. One option is a deal where Nvidia hires Reflection's staff and licenses its technology instead of buying the company outright, which could avoid a long antitrust review. An agreement could come within weeks, the paper said, though the talks could also fall apart. Nvidia has already put about $800 million into Reflection, which was raising money in April at a pre-money valuation of $25 billion. Neither company commented. Reflection has said Beam's weights will be released later this month.

Eleven tech companies pledged a combined $2.4 billion in computing credits and model access for the Genesis Mission, the federal push to use AI for scientific research, at a White House science summit on October 8. Nvidia committed $1 billion, AMD $500 million and OpenAI $200 million, while Anthropic and Google pledged $150 million each. Google had already given $40 million in July. The credits will support research at more than 15 federal agencies, including NASA, the National Institutes of Health and the National Science Foundation. Anthropic said its share covers Claude, Claude Code and API credits for several hundred projects over three years, with fusion energy and quantum computing among its priorities. The administration billed the event as a $6 billion push in all.

Microsoft released Microsoft-Decision-1 on October 9, joining OpenAI and TypeSafe in the fast-growing category of decision models. Instead of generating an answer, it takes a fixed set of choices and returns a calibrated probability for each one, which software can act on directly. Microsoft built it by further training Alibaba's open Qwen3.5 model with 9 billion parameters, and says it plans to rebuild it later on Microsoft and OpenAI models. The company says it beat both chat models and rival decision models on accuracy across 36 benchmarks, and ran 35 times faster than GPT-6 Sol. It aims the model at routing, classification, grading and agent safety checks. It is available now in Microsoft Foundry and through OpenRouter.

An investigation by KFF Health News found that the Centers for Medicare and Medicaid Services has run a Slack workspace since August 2025 where tech companies and investors talk directly with federal health officials about AI health apps and access to patients' medical records. It has grown to about 1,700 members, with only a handful of patient advocates, doctors and hospitals among them. In February, officials used it to invite companies including Microsoft, Anthropic, OpenAI, Apple and Google to an unannounced Food and Drug Administration session on chatbots for patients. Legal experts told KFF the group resembles a federal advisory panel, which would have to meet in public. The American Medical Association's chief executive said some of these tools "just aren't ready for primetime."