Anthropic introduces a new standard aimed at simplifying connections between AI and machines.

      The Model Hardware Standard (MHS) offers a universal translation layer, enabling AI agents to operate laboratory instruments and manufacturing equipment with standardized controls.

      Anthropic has released a research preview of its Model Hardware Standard, a unified specification aimed at allowing AI agents to safely manage lab tools and factory machinery. The company asserts that MHS significantly reduces the integration time for complex machinery with AI, shrinking it from months to mere hours or minutes. Currently, the standard is accessible to select research labs and manufacturers via a waitlist, with plans to make it open-source in the future.

      **Connecting AI to Physical Devices**

      Similar to Anthropic's Model Context Protocol (MCP) that established software standards, MHS functions as a universal translator for physical hardware. It eliminates the need for custom software for each instrument, allowing devices to communicate through standardized commands and automatically generated reference documents that outline operational parameters and safety constraints.

      The system is model-agnostic and compatible with any hardware that has a programmable interface. Early trials in real-world applications have shown substantial efficiency improvements across different research environments:

      - **QuEra Computing**: A team of four engineers previously took months to create a laser-relocking script, which had a recovery time of 150 seconds and a success rate of 58%. After running an overnight optimization loop, four Claude instances revamped the process, reducing recovery time to just six seconds with a development success rate of 96% and 99.3% in a 700-trial blind test. Claude also optimized 12 interdependent servo parameters over 16 unattended hours, lowering the residual error from 15.7 mV to 1.55 mV.

      - **Carnegie Mellon University**: Researchers managed to integrate a liquid handler, plate reader, robotic arm, and cameras across three computers in just eight hours, a process that usually takes weeks. With the assistance of a Claude Opus 4.8 agent, the system executed serial dilution experiments three times faster. In safety tests, MHS effectively prevented six induced fault conditions before any hardware initiated movement.

      - **Genentech**: Engineers implemented MHS in an automated protein assay workflow, allowing Claude to autonomously optimize liquid transfer rates for different viscosities.

      **Initial Challenges and Future Directions**

      Despite the achievements in the initial preview, certain operational limitations have emerged. For instance, at Genentech, Claude repeatedly tried to resolve fluid-handling bubbles by executing the same software command, necessitating human intervention to clarify that the problem was physical rather than coding-related. Additionally, since AI models predominantly interpret physical environments through text logs and images, spatial reasoning still needs expert human supervision.

      Anthropic is collaborating with hardware partners such as Universal Robots, Tecan, and AWS to broaden MHS support. Integrations are also progressing towards consumer-adjacent platforms; Hugging Face is planning to support its LeRobot project, and Raspberry Pi is testing driver compatibility. Anthropic has not yet disclosed a definitive date for the open-source release, public schema, or specific governance framework for the standard.

      Pranob is an experienced tech journalist with over eight years in the consumer technology field, where he has contributed extensively.

      South Korea is providing free AI services to assist citizens with appointment bookings, tax filings, and other everyday tasks. While discussions on AI regulation continue globally, South Korea aims to make access to AI as straightforward as electricity or public transportation. The government intends to finance premium AI tools, allowing all citizens to benefit through a new initiative known as 'AI for All.' As reported by the Wall Street Journal, the program aims to enhance the usage of domestic AI chatbots while decreasing reliance on AI models from the US and China.

      OpenAI and more than 100 other companies are sounding the alarm about a potential surge in AI-driven cyberattacks. They warn that as AI capabilities grow, these sophisticated attacks could become more common and harder to counter. The coalition is advocating for a global initiative focusing on bolstering cyber defenses, as essential systems such as hospitals, water treatment facilities, and internet infrastructures face increased risk if defenses do not improve swiftly.

      Gemini Notebook, previously known as NotebookLM, is introducing a new feature that allows users to reference books they own and convert them into varying formats. Users must have purchased the books via Google Play Books to add them as sources in Gemini Notebook. Google refers to this feature as Expert Intelligence, which enables users to extract knowledge from their digital books instead of relying solely on uploaded sources or internet content. Currently, Google is offering one free book to users in the US as part of the launch, while supplies last. Expert Intelligence is initially available in the Gemini Notebook app and web dashboard, with plans to integrate it into the main Gemini app and AI mode in Search in the future.