Anthropic Gives AI a License to Touch Hardware: MHS Lets Agents Operate Machines Safely
After learning to type and learning to code, the next step for large models is learning to use their hands — and the prerequisite for using hands is knowing what not to touch.
Three Key Facts
Anthropic released a research preview of its Model Hardware Standard (MHS) this week, aiming to build a "universal language" between AI agents and physical devices: lab instruments and manufacturing machines from different vendors can be connected, read, and operated by agents through a standardized interface. Anthropic says that wiring up devices from different brands into one automated workflow currently takes weeks or months of custom integration; MHS aims to cut that to hours or even minutes.
Safety design is the core selling point. Each device's driver file declares its capabilities, parameters, physical characteristics, and — most importantly — its safety limits. What an agent reads isn't just "what this machine can do," but "where it must stop." Anthropic says early tests show MHS halting an operation before a device moves into an induced unsafe state. The standard is model-agnostic and runs over open protocols like the Model Context Protocol (MCP).
MHS is currently open as a research preview to partners in science, robotics, electronics, and manufacturing. Anthropic plans to eventually open-source it, along with a physical safety roadmap and adoption guidance.
WangDou's Take
The interesting part of this story is that Anthropic didn't open with "AI can run your lab" — it spent most of its airtime on "when must AI stop." For two years the agent industry's main narrative has been "how much can it do"; MHS flips the question to "do we dare let it." Because the physical world isn't a screen: there is no Ctrl+Z. A fried instrument, a spilled sample, an injured worker — none of those can be fixed by rolling back one line of code. Whoever turns "safety limits" into a standard first gets the ticket to physical AI; whoever only flexes automation is burying landmines for the whole industry. Before giving machines hands, write down the rules — that's what actually being ready for AI on the shop floor looks like.
Source: Computerworld, PYMNTS