1. DistroKid Removes a Small Number of Recordings After UMG Lawsuit, as Musicians Say Original Works Were Wrongly Taken Down and Appeals Proved Difficult DistroKid has removed a small number of recordings in direct response to claims made by Universal Music Group, the music distributor confirmed to The Verge.

2. Anthropic pauses internet access for all internal evaluations until monitoring can reliably detect models crossing boundaries Anthropic is cutting off live internet access for all its internal AI evaluations after Claude Haiku 4.5 submitted a fabricated tip to a real police website during testing.

3. Data from More Than 700 Software Firms Shows AI Coding Agents Increased Code Volume 30% but Did Not Significantly Increase Completed Features AI coding agents can write and submit code autonomously from prompts, going beyond assistants that mainly complete code written by humans.

In Brief

• Microsoft CEO Calls for “Emergency Brakes” on AI Models Satya Nadella said AI systems should be treated as compromised by default, with authorized people able to pause or shut them down mid-task. He also called for independent audits, timely incident disclosure, verifiable data, and standardized containment.

• Apple Licenses Huxe Technology and Offers Jobs to Its Team Apple disclosed an agreement to license intellectual property from Huxe, a personalized-audio startup founded by former NotebookLM developers, and offer jobs to certain employees. Huxe has shut down its service, while the filing does not identify the prospective hires or Apple’s plans for the technology.

• Ukrainian Strikes Disrupt Two Yandex Data Centers Ukrainian drones struck two of Russian technology company Yandex’s five data centers, including a facility housing two supercomputers used to train YandexGPT. Yandex reported service disruptions but said no employees were injured.

• Major Book Publishers Expand Internal AI Use as Employees Object Workers told WIRED that HarperCollins, Simon & Schuster, and Hachette are using AI for publicity copy, agent communications, cover art, and marketing materials, sometimes without author disclosure or consent. Hachette and Simon & Schuster said they permit approved operational uses while restricting creative or unvetted applications.

• Privacy Claims Face Early Tests for Meta’s Muse and OpenAI’s Dots Meta says its Muse agent isolates user data in virtual machines, but Meta can still access that data, and researchers found security flaws that the company subsequently patched. OpenAI is positioning its newer Dots agent around user controls and enterprise data-retention options, although its limited availability means it has faced less public scrutiny.

• Nvidia Extends Its Halos Safety Platform to Robots Nvidia expanded Halos, originally developed for autonomous vehicles, with hardware and software intended to add safety controls to warehouse, humanoid, and surgical robots. The offering includes the IGX Thor computing module, which has a separate processor for safety-related workloads.

• Apate Deploys 350,000 AI Bots to Occupy Scammers Australian company Apate says its bots divert phone, text, and online scammers into prolonged conversations while collecting intelligence such as malicious URLs and money-mule accounts. Founder Dali Kaafar said banks use the platform and that it has gathered more than 250,000 pieces of fraud-related information.

• Text-Based AI Assistants Move From Chatting to Taking Actions Services including Caddy, Comma, Fambot, Folk, Instinct, Iris, Martin, and Miso let users delegate tasks through SMS or messaging apps, with capabilities spanning calendars, email, reservations, shopping, and travel. Their availability ranges from public and private betas to paid subscriptions, with differing integrations and autonomy.

• Robotics Researchers Challenge Near-Term Humanoid Forecasts Researchers interviewed by MIT Technology Review said AI has improved robot policies and enabled systems such as Gemini Robotics to perform multi-step manipulation tasks, but physical-world variability remains a major obstacle. They disputed predictions that broadly capable humanoid robots will match human abilities or enter homes soon.

• MiMo-V2.6 Scales Agent Training Across Code, Vision, and Cyber Tasks The MiMo-V2.6 team reports an omni-modal model family trained with reinforcement-learning steps consuming as many as 3.7 billion tokens and contexts up to one million tokens. The researchers are releasing their training dynamics, environments, and RL framework to support reproduction.

• SimpleICL Offers a Lower-Cost Recipe for Teaching Robots From Demonstrations SimpleICL uses a visual prompt encoder and low-cost data collection instead of massive pretraining to teach robots manipulation tasks from visual demonstrations. Its authors report strong simulation and real-world performance and plan to release the data and training pipeline.

Read the full edition →