11 October 2026

Nadella: treat AI models like an insider threat and build an "emergency brake"

Microsoft CEO Satya Nadella published a long post on X about controlling advanced AI models. He proposes separating the model from the system that controls it and assuming from the start that the model is compromised, meaning it may act against us, so it has to be surrounded with deterministic safeguards and procedures. An authorized person should always be able to stop the model mid-task, "like an emergency brake." Every significant action by the model should leave a tamper-resistant, human-readable trail, and transparency of reasoning is, in his view, "non-negotiable." "The most trustworthy superintelligence system will not be the one in which we trust the model the most, but the one that lets us trust it the least," Nadella writes.

Satya Nadella na X ↗ CNBC ↗ The Verge ↗

From recent days

New York court sentences Michael Smith to 18 months in prison - his bots played AI songs billions of times, earning him over $8 million in royalties

Judge John G. Koeltl on Tuesday, October 6, sentenced Michael Smith, 54, of North Carolina to 18 months in prison and two years of supervised release and ordered the forfeiture of $8,091,843.64, the US Attorney's Office for the Southern District of New York said. According to prosecutors, from 2017 to 2024 Smith ran thousands of fake accounts (up to 10,000 at once) that streamed hundreds of thousands of AI-generated songs billions of times; in April 2023, Smith's bots using YouTube Music family plans played his music 80.9 million times, while Taylor Swift's entire catalog had 9.3 million plays on such plans. Smith pleaded guilty in March to conspiracy to commit fraud; it is the first criminal streaming fraud case in the US, Billboard reports, adding that the defense asked for no prison time and prosecutors for at least 46 months.

justice.gov ↗ Billboard ↗ Gizmodo ↗

Anthropic releases Claude Haiku 5.5 - the company's fastest model, about 75 percent cheaper than its predecessor on average

Anthropic released Claude Haiku 5.5, a small model in the Claude 5.5 family for high-volume, low-cost tasks: summaries, classification, database queries and working as a coding subagent. According to the company, it is its fastest model ever, and using it costs about 75 percent less than Haiku 4.5 on average: $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. It is the first Haiku with adjustable reasoning effort. Anthropic also halved the price of cache reads for Sonnet 5.5 and is introducing a monthly API credit: $100 for Max 5x subscribers, $200 for Max 20x and up to $500 for Team.

Anthropic ↗ The Verge ↗ 9to5Mac ↗

Anthropic's IPO prospectus: $42 billion net loss in 2025 and a warning of "existential" AI risk

Anthropic devoted 80 of the 261 pages of its prospectus to risks and warns in it that advanced AI may pose "catastrophic or existential threats to humanity," Reuters reports. The document shows that the company's revenue grew twelvefold in 2025, to almost $4.6 billion, and its net loss was $42 billion. At the valuation above $2 trillion that investors are hoping for, it would be the largest stock market debut in history.

The Verge ↗ TechCrunch ↗ The Verge ↗

All news →