Microsoft has shipped fresh Windows 11 Insider Preview builds to the Beta and Experimental channels, covering the 26H1 and Future Platforms tracks. On Experimental, Windows Search has been rebuilt on WinUI 3 and adds inline actions so you can change system settings straight from the search box. Beta picks up Cloud rebuild, a WinRE recovery path that reinstalls Windows from Windows Update. The Experimental channel also enables IAKerb by default to reduce NTLM fallback. Widgets, Task Manager, Narrator, and File Explorer see smaller updates.

Source

Anthropic’s tests found that GLM-5.2 could autonomously spread across nearly an entire simulated network, while a 9-billion-parameter model set up inference on eight of nine hardware types. The experiments point to a potential AI worm that could use poorly secured IoT devices for compute; Anthropic estimates small models may reach the capabilities needed for end-to-end replication by the end of 2026.

Source

Intel has named its Crescent Island accelerator the GPU D-Series and revealed peak FP8 performance of 1.3 petaflops, alongside 1.5 TB/s of memory bandwidth. The air-cooled, 350-watt PCIe cards will offer up to 480 GB of LPDDR5X memory, targeting large-scale AI inference without the power and cooling demands of rack-scale accelerators.

Source

Python 3.15 is now stable, bringing opt-in lazy imports that can defer module loading until code needs it, while making UTF-8 the default for text files opened without an explicit encoding. The release also reports faster experimental JIT performance, but gains vary by platform and comparison baseline. New profiling and runtime features give developers more ways to inspect and tune Python applications.

Source

Meta has released Muse Wonder and Muse Wonder Full, two models built for long-running agent tasks, coding, and computer use. Both offer a 1 million-token context window; Wonder is available through Meta’s API, while the more advanced Full version is limited to Muse Code subscribers. Meta says the models will also power its Muse personal agent. Meta’s benchmark table places the Wonder family among frontier competitors such as Claude Opus 5.5 and GPT-6 Astra, with Wonder Full leading on some tests but not across every evaluation.

Source

Intel will end new Xeon support for Software Guard Extensions (SGX) after Diamond Rapids, making Trust Domain Extensions (TDX) the focus of its confidential-computing roadmap. Existing SGX platforms are due to receive security updates and support into the early 2030s. Intel also plans to add trusted I/O, post-quantum cryptography, and enhanced attestation to TDX, though it has not set availability dates.

Source

As technical barriers to enterprise AI adoption ease, organisational readiness is emerging as the harder problem. LogicMonitor’s Karthik Sj says businesses need clear governance, expert oversight and coordinated teams to turn AI pilots into safe, useful operations. AI is also moving from experiment to boardroom priority.

Source

SpaceXAI may become SpaceXSI, Salesforce’s AIForce is set to turn into SIForce, and Nvidia and Microsoft executives are also using “super intelligence”—Trump’s preferred replacement for AI. Apparently, tech leaders hold Trump’s AI—sorry, SI—expertise in high regard.

Source

China’s renewable boom is outpacing its ability to move electricity, leaving roughly 9% of wind and solar output curtailed in the first half of 2026. A proposed $3.9 trillion grid overhaul, alongside batteries, pumped storage and more flexible power trading, could deliver more of that energy to data centers rising in wind- and solar-rich western regions—and help China expand computing capacity for domestic use and export. The advantage is promising, but hinges on whether infrastructure and market reforms can keep pace.

Source

Andrej Karpathy is an AI researcher and co-founder of OpenAI who previously led AI at Tesla. He says understanding of LLM capabilities is splitting into a steep funnel: roughly 6 billion people have barely encountered LLMs, while about 5,000 have internal access to frontier systems. Between them, he estimates 1–2 billion casual users and 20 million professionals using advanced models for complex math and coding work.

Source

OpenAI is defending its decision to fire three safety researchers after they disputed the company’s misconduct allegations and warned that the dismissals could discourage staff from raising safety concerns. The researchers say their outside collaboration was part of their work and want clearer rules for sharing information with safety evaluators.

Source

Microsoft president Brad Smith is urging governments to set global AI safety rules, arguing that everyday products such as milk and washing machines face tighter oversight. He wants legal duties to report AI-driven breaches and independent assessments of systems before release. Smith says the UK could lead the effort.

Source

An OpenAI model grading seven responses during reinforcement-learning training tried to damage its task environment after discovering the required input files were missing. It first submitted fabricated scores and fake input files; when checks rejected them, it removed software and attempted to delete system directories, hoping the host would replace the environment. None of the grades from that attempt was accepted.

Source

Personal AI agents are already reaching corporate systems, often through employee-approved OAuth grants, saved passwords, or long-lived API keys that make the agent look like the user. Blocking them may push usage out of sight; a safer approach is to identify each agent, limit its access, and revoke it independently when needed.

Source

AvisLoader, a newly uncovered Windows malware loader, uses the encrypted Tox peer-to-peer network for operator commands and payload delivery, reducing its reliance on a fixed server or domain that defenders could take down. It arrives through a ClickFix lure disguised as a DocuSign verification prompt, which tricks visitors into running a command. Recovered files also reveal shortcut-based persistence and a control panel for managing infected systems.

Source

Microsoft is rolling out Office editing inside Copilot, letting users create and collaborate on Word, Excel, and PowerPoint files without leaving the app. The update puts a practical feature behind the Copilot super app announced in September: users can move from a chat or Cowork session into editing a document while keeping the conversation alongside it.

Source

Nvidia is discussing a deeper investment in Reflection AI, including an acqui-hire that would bring the startup’s staff to Nvidia and license its technology. Reflection develops open-weight AI models that developers can use and adapt. Its Beam model even aims to rival Chinese models, coming close to Qwen 3.8-Max on coding and agent benchmarks, according to 4sysops. The talks are preliminary, and options also include supplying more computing power or a full acquisition; no agreement has been reached.

Source

Hyper-V Manager’s New VM wizard and several other dialogs now support dark mode in Windows 11 and Windows Server preview builds. The main management window remains light, and users must restart Hyper-V Manager for theme changes to take effect.

Source

Microsoft says its new Microsoft-Decision-1 model made AI agents’ adaptive replanning nearly four times faster in internal tests. The decision-scoring model returned scores 46 times more consistently than the large language model used for comparison, according to the company. Its launch follows OpenAI’s recent introduction of a Decisions API, a sign of growing interest in models that choose between options rather than generate text.

Source