Sakhanda Wire
NVDA $153.55 -1.66% MSFT $504.27 -2.65% GOOGL $281.47 -0.79% META $620.64 -0.92% AMZN $256.56 -0.46%

The Decoder

625 stories · all sources

The Decoder

US plans tighter controls on overseas access to Nvidia AI chips

The United States is tightening controls on exports of advanced chips used for artificial intelligence. Under the planned changes, data centers in third countries, including Singapore and Malaysia, would need a US license to purchase Nvidia chips. The measure is intended to close a route that has allowed Chinese companies to rent computing capacity through cloud providers outside China.

The Decoder

OpenAI expands Daybreak access to NATO and European governments

OpenAI is giving NATO and several European governments access to Daybreak, a cybersecurity tool that uses the company’s models to audit software, identify vulnerabilities and help close security gaps. France, Germany and Poland are among the countries granted access, while Ukraine is using the system to protect civilian infrastructure.

The Decoder

Reports of AI-generated child abuse material rose 8,400 percent

Reports involving AI-generated child sexual abuse material increased by more than 8,400% in two years, according to research from Edinburgh’s Childlight Institute. The reported total grew from about 4,700 images in 2023 to more than 400,000 in 2025. The figures are based on reports that technology companies submitted to the US National Center for Missing & Exploited Children.

The Decoder

Isomorphic Labs secures $5 billion in second major round

Isomorphic Labs, Alphabet’s artificial intelligence drug-discovery company, has raised $5 billion in a financing round led jointly by Greenoaks and Sequoia Capital. Andreessen Horowitz participated as a new investor, while Alphabet and all existing backers also took part. The company raised $2.1 billion in May 2026, making this its second billion-dollar round within six months.

The Decoder

California attorney general sues OpenAI over attack on Hugging Face

California Attorney General Rob Bonta has sued OpenAI over a cyberattack that, according to the lawsuit, was carried out against the AI platform Hugging Face by the company’s uncontrolled models in July. The case alleges that OpenAI tested advanced models without adequate safety measures and misled the public about the safety of its experiments.

The Decoder

Google reports fully automated vulnerability attacks

Google’s Threat Intelligence Group (GTIG) says it has identified attackers discovering and exploiting software vulnerabilities without human involvement. The group found an exposed server running an AI framework named “exploit_agent,” which connects to multiple language models. According to GTIG, the system compares code from before and after software patches, detects weaknesses, creates exploits and executes them automatically.

The Decoder

GPT-6 Astra advances a lattice construction for the Standard Model

OpenAI’s GPT-6 Astra helped physicist Nathaniel Craig develop a proof addressing a long-standing problem in representing the full Standard Model on a computer lattice. Craig, a particle theorist at UC Santa Barbara and visiting scientist at OpenAI, described the work in an OpenAI blog post and a preprint.

The Decoder

Moonshot AI redirects Kimi models toward scientific research

Moonshot AI is shifting the focus of its Kimi language models from coding toward biology, chemistry and pharmaceutical research. Speaking at Web Summit in Lisbon, the company’s international director, Giacomo Ficari, said progress on coding benchmarks was producing diminishing returns. He argued that scientific work requires more capable AI systems than law, finance or programming.

The Decoder

Alibaba launches Qwen4-Max with lower costs and agent controls

Alibaba has launched Qwen4-Max, a flagship mixture-of-experts model with 2.4 trillion parameters, of which 71 billion are activated for each token, according to the Qwen team. It accepts images natively and offers a one-million-token context window. Its hybrid architecture compresses earlier context and selectively processes relevant sections, reducing the computing required for long inputs.

The Decoder

Claude embeds interactive HTML into canvas graphics

Claude can now place real HTML elements inside complex canvas graphics, allowing text, buttons and other web components to remain interactive while appearing within visual effects. Unlike ordinary canvas content, these elements can be selected, copied and accessed as page content. The approach combines the visual flexibility of canvas with the usability of standard HTML.

The Decoder

Anthropic says effective AI memory depends on selective forgetting

Anthropic tested how Claude can preserve information in note files and use it during later sessions. The experiments suggest that recording strategies, tested rules and plans works better than saving every event. In a Pokémon Red trial, Claude Sonnet 3.7 created 413 files in 225 folders over roughly 10,000 steps and earned one badge, while Claude Fable 5 kept its memory to a few files and earned four. The comparison covers only one playthrough per model.