Industry news

Microsoft researchers say they have developed a cost-effective machine learning method that improves weather forecasts two to six weeks ahead, where physics-based and existing AI models typically lose reliability beyond two weeks. The findings were published in the journal Nature, and the tool is now open-source software.

Square is launching the Square Kiosk, a hardware device with a 21.5-inch touchscreen and a large pill-shaped design. It includes an agentic, conversational voice layer that understands a seller's menu and responds in real time, letting customers ask questions and request modifications or substitutions. Square says the device can greet customers as they approach using presence detection.

Android is adding scam defense features, including an "enhanced protection" setting for Google Messages that will roll out in the coming weeks. The setting uses cloud-based Gemini AI models to evaluate a possible scammer's intentions in conversations with unrecognized senders.

A Saudi national who worked as an engineer at OpenAI was arrested by the FBI and charged with using or attempting to use a device to intercept oral communications. According to the complaint, he planted a recording device hidden in toilet paper behind a television in a conference room used for confidential discussions. OpenAI said it found no indication that he accessed or exfiltrated model weights, source code or customer data.

Anthropic's blog post warns that small AI models could be "months away" from spreading between smart household devices such as fridges, ovens, vacuum cleaners, security cameras and TVs. The article frames these self-replicating "AI worms" as either an imminent cybersecurity risk or overstated hype, depending on the observer.

Picus Security researcher Sila Ozeren Hacioglu argues that agentic pentesting, where AI agents run tests step by step, finds complex exploit chains faster than humans but leaves blind spots. The article says safety restrictions and network limitations keep many assets and vulnerabilities untested, and that combining agentic pentesting with exposure validation and breach simulation can help close those gaps. It notes that agentic runs are costly and vary between runs, and that Omdia research found organizations pentest only 32% of their attack surface on average.

Socket for Origin is now available in the Origin App Marketplace. It scans supported manifests and lockfiles when a pull request is opened or new commits are pushed, and posts a policy-based security check and dependency findings directly on the PR to help teams catch risky packages and compromised updates before they merge.

Amp, a developer-focused AI agent company, describes using OpenAI models for most of its token usage, including GPT-6 Astra, and new integrations such as Sign in with ChatGPT, Puck powered by GPT-Live-1, and Ultrafast mode. The article presents these as ways to let agents take on larger tasks beyond writing code, such as planning, writing, and design.

SentinelOne reports that OpenAI Daybreak, which gives defenders access to OpenAI models for cybersecurity work, reduces malware reverse engineering from four to six weeks to roughly eight hours, according to Chief AI Officer Gregor Stewart. The faster analysis lets SentinelOne researchers examine more samples and spot shared components and relationships between threats. Stewart says the company still keeps humans accountable for architecture decisions, while smaller fixes may suit automation.

Voice clones and synthetic video calls are now appearing in payment fraud, help desk resets and access requests. The article argues that a structured readiness drill can show security teams where verification breaks down without any real risk. It cites a 2025 Gartner survey of 302 cybersecurity leaders in which 62% reported at least one deepfake attack in the prior 12 months.

XBOW researchers chained three previously unknown Google Chrome bugs into a full-chain exploit that runs code on a Windows computer when a user only loads a web page. They built it in six days by directing AI agents running OpenAI's publicly available GPT-5.6 Sol model, and Google fixed all three bugs and paid XBOW a $250,000 reward.

Wiz has made its Wiz AI Agents available within Wiz for U.S. Government (Wiz for Gov), a FedRAMP High authorized environment, positioning it as the first AI-APP for U.S. public sector organizations. The agents extend CNAPP with AI-accelerated insights, and security teams can configure whether they remediate routine issues autonomously or require human-in-the-loop approval.

Nvidia is donating $250m worth of computing power to the UK over five years to support AI safety research. The UK government confirmed that its AI Security Institute will be able to examine cutting-edge models from companies such as OpenAI, Anthropic and Google before they are released.

US law enforcement warns that attackers are using generative AI chatbots to plan attacks, citing a May 2025 car bombing outside a Palm Springs fertility clinic. A law enforcement memo reviewed by the Guardian says the suspect used an unnamed generative AI tool that failed to detect his bomb-making request, and that the conversation's privacy hindered police alerts.

The heads of the Bank of England and the International Monetary Fund warned that rising cyber-attacks risk undermining trust in the financial system. Speaking on a panel about online fraud at the IMF and World Bank meetings in Bangkok, Andrew Bailey pointed to headlines about self-replicating AI agents hijacking personal devices and a French supermarket hack that heated fridges up to 50C.

Third Moment Research is building a self-improving AI system that uses OpenAI models, including GPT-6 Astra, GPT-6.1 Sol and Codex, to run 10,000 agents at once on investigations lasting 36-48 hours. The company also uses Daybreak to test the security of the environments where its agents work. In one quantitative finance case, Codex cut a trading strategy backtest from approximately 3 hours to approximately 3 to 5 minutes.