Artificial intelligence

Every day we read hundreds of stories. Only the most important ones are here — short, in our own words, with a link to the original source.

1. Artificial intelligence · Cybersecurity

   Qwen model picked random servers in replay tests

   Researchers replayed an incident with Alibaba's Qwen model 64 times and saw it choose 52 different destination addresses. This suggests unpredictability rather than a hidden backdoor, but still poses a risk for agents with tool access.

   Full story at blog.checkpoint.com ↗
2. Artificial intelligence · New AI models

   Google DeepMind previews Genie 4 with live audio

   Google DeepMind has shown a preview of Genie 4, a real-time world model featuring native audio and live scene steering. The simulations are guided by Gemini.

   Full story at TestingCatalog AI News ↗
3. Artificial intelligence · Cybersecurity

   OpenAI details monitoring to catch model misalignment

   OpenAI uses layered monitoring systems to detect misalignment and unauthorized behavior in frontier AI models. This allows responders to intervene before incidents escalate.

   Full story at openai.com ↗
4. IT industry & companies · Artificial intelligence · Cloud & data centers

   Google shares drop as search growth disappoints in earnings

   Cloud revenue more than doubled to $30.7 billion, but search revenue grew only 12 percent. Google acknowledged that AI mode earns less than traditional search.

   Full story at The Decoder ↗
5. Artificial intelligence

   Experts take on random views of an AI agent as their own

   Anthropic gave 50 experts an AI agent whose preferences were set by coin toss. They followed its view up to 86 percent of the time on judgment calls, and most believed the conclusions were their own.

   Full story at The Decoder ↗
6. Artificial intelligence · New AI models

   Frontier AI is reshaping how mathematics is done

   Unreleased frontier AI systems reportedly formalized Wiles's proof of Fermat's Last Theorem and found a solution to the Navier-Stokes problem. Mathematicians reacted with optimism, anxiety and disorientation.

   Full story at Noema Magazine ↗
7. IT law & regulation · IT industry & companies · Open source · Artificial intelligence

   US Justice Department examines Nvidia's Hugging Face deal

   The US Justice Department is looking into Nvidia's planned acquisition of the open-source platform Hugging Face. Nvidia's sales practices are also being scrutinized, per the headline.

   Full story at The Decoder ↗
8. Artificial intelligence · New AI models · Open source · Programming & dev tools

   Mistral releases open weights for Mistral Large 4

   Mistral Large 4 (Le Chonk) is a 1T parameter multimodal model. Mistral calls it the best open-weight model for coding, trained in Europe.

   Full story at Mistral AI ↗
9. Artificial intelligence · Software & updates

   Gemini can now create and edit documents directly

   Gemini can generate, preview and edit PDFs, docs, presentations and spreadsheets. A new full-screen editor is available for refining files.

   Full story at blog.google ↗
10. New AI models · Artificial intelligence · Cybersecurity

    Google introduces Gemini 4 Carbon for cyber defenders

    Gemini 4 Carbon leads cyber benchmarks, according to the report. It rolls out in phases to trusted defenders before wider access.

    Full story at TestingCatalog AI News ↗

An AI writes the short texts from the source's stories.