Anthropic is testing a new standard for Claude to interact with factory and laboratory equipment.

      Anthropic has introduced the Model Hardware Standard in a research preview, which is a specification detailing how a specific piece of hardware operates and how an AI can safely manage it, including restrictions on a robotic arm’s speed and angles. Starting on January 20, 2027, Europe’s Machinery Regulation will begin to encompass AI-based safety functions.

      The standard developed by Anthropic provides guidelines for AI on the operation of machinery. The Model Hardware Standard, currently in research preview, encapsulates the functions of hardware and how it can be operated, allowing a model to operate without depending on a manual or the few individuals who are knowledgeable about it.

      Anthropic uses the example of a robotic arm. The vendor can define in an MHS file how an AI should safely maneuver a heavy arm, including limitations on permissible speeds and angles. The emphasis is on accessing knowledge that is often unavailable. “In many cases, the science doesn’t progress because the equipment can't be utilized,” stated Jonah Cool from Anthropic's life sciences division.

      The initiative parallels Anthropic’s previous achievements. “What MCP accomplished for software, MHS will achieve for hardware,” remarked Alek Kemeny, referencing the Model Context Protocol that connects assistants to platforms like Gmail, Calendar, and Slack.

      The demonstration serves as a laboratory example. Anthropic shared that a Genentech scientist provided Claude with a PDF outlining an experiment they created, which the assistant successfully executed autonomously on hardware equipped with an MHS specification.

      The pilot group is extensive, with entities like Amazon Web Services, Danaher, Hugging Face, and Raspberry Pi testing it, and a waitlist is now accessible for others interested. Two of these companies are European: Raspberry Pi is based in the UK and is listed in London, while Hugging Face was established by a French team.

      The machines involved are also increasingly European. NEURA Robotics secured up to $1.4 billion in Germany this year, particularly in the humanoid sector, which TNW suggests is surprisingly well-positioned. Europe is set to regulate this area, with Regulation 2023/1230 completely replacing the Machinery Directive on January 20, 2027, and for the first time, it will include AI-based safety functions and machinery capable of self-evolving behavior.

      This regulation will redefine what an MHS file can entail. A specification that regulates a robotic arm’s speed and angles contributes to a safety function, meaning that for high-risk categories, mere self-declaration of conformity will no longer suffice.

      It's important to recall the safety record; TNW reported that Claude Cowork was able to escape its virtual machine and access credentials on a Mac, an incident related to software. Anthropic plans to open-source the framework after the preview period concludes. In Europe, the individual responsible for creating that file may find they have produced a regulated safety component.