Anthropic's Bet: One Safety Spec for Every AI Robot

On August 27, Anthropic put out something that got less attention than it deserved: a specification called the Model Hardware Standard, built to let AI agents run microscopes, liquid handlers and robotic arms through one common interface. The pitch underneath it is bigger than lab automation. Anthropic is trying to write the safety rules for physical AI now, while language models still mostly live in chat windows, rather than after they are wired into machines that can crush a hand. Whether OpenAI, Google and the robotics companies pick it up will say a lot about how the next few years go.

What Is the Model Hardware Standard?

The Model Hardware Standard (MHS) is an open specification, developed with the HHMI Janelia Research Campus, that stretches Anthropic's Model Context Protocol out of software and into hardware. MCP gave models a standard way to call tools, APIs and data sources. MHS does the same job for physical devices: an agent can discover a machine, read its sensors and write to its actuators without an engineer hand-coding a custom integration for every model of pipetting robot.

The problem it targets is real. Lab and factory equipment ships with proprietary drivers, odd APIs, and operating knowledge that lives in a PDF manual or in one technician's head. Anthropic says connecting a new device drops from weeks or months to hours or minutes. "In many cases, the science doesn't happen because you can't use the equipment," said Anthropic's Jonah Cool.

How the driver actually works

Each device gets a standardized driver that translates between the agent and the hardware using two basic commands: read and write. Next to it sits a reference file describing the machine — what it can measure, which parameters can change, how much a part weighs, and the limits it has to respect. A robot-arm vendor can pin down the maximum speeds and joint angles an agent is allowed to use. A microscope vendor can cap laser power so a model cannot cook the sample. The agent reads those constraints before it moves anything.

Why Anthropic Wants the Rules Set Now

Timing is the whole story. Right now the number of LLM-driven systems touching real actuators is small and mostly penned inside supervised labs. That window is closing fast. Figure, Tesla, Google DeepMind and a long list of startups are pushing general-purpose robots, and the control layer is increasingly a foundation model. Anthropic's bet is that a safety convention adopted while the field is small becomes the default once it is large. Set it late and everyone is retrofitting.

Safety baked into the spec

• Device-level limits the agent cannot override: collision boundaries for arm transfers, power ceilings for lasers, allowed speed and angle ranges.
• State checks before an action runs, such as detecting a missing or rotated sample plate.
• Emergency stops that fire on anomalies or when hardware stops responding.
• Explicit rules for when a human has to approve a step instead of letting the agent proceed on its own.
• A per-device reference file so the agent works from stated physical constraints rather than guessing.

"Claude learns about the physical world through text and images, meaning its spatial and physical reasoning have limitations that still require expert oversight," Anthropic says. In one Genentech test the model tuned a workflow but missed that its own liquid handling was making the samples foam.

Will Rivals Actually Adopt It?

The optimistic case is MCP itself. Anthropic released it in November 2024, and within months OpenAI, Google DeepMind, Microsoft and Salesforce had all adopted it. There are now more than 16,000 MCP servers in the wild, and Anthropic has since handed the protocol to a Linux Foundation body co-founded with OpenAI and Block. The New Stack called it one of the fastest protocol adoptions in tech history. A neutral standard with no obvious lock-in spread because fighting it cost more than joining it.

Hardware is a harder sell. The early roster leans toward vendors and labs — AWS, Danaher, Doosan Robotics, QIAGEN, Tecan, Universal Robots, plus Hugging Face and Raspberry Pi — not competing model labs. The spec also still sits behind a waitlist and stays proprietary through the preview, with open-sourcing promised only after Anthropic finishes building safety evaluations with its partners. If a robotics company is sprinting to ship a humanoid, a standard that adds approval gates and speed caps reads as friction. That is the real test: not whether the idea is good, but whether safety-first survives contact with a hardware race where the payoff goes to whoever moves first.

The regulatory clock

There is a forcing function. The EU's Machinery Regulation (2023/1230) takes effect on January 20, 2027, and for the first time it covers AI-based safety functions. An MHS file that constrains how a machine moves could count as a regulated safety component, which means formal third-party assessment rather than a vendor ticking a self-certification box. That cuts both ways. It gives companies a reason to standardize on something common and well documented, and it drops real legal weight on whoever ends up authoring those safety specs.

Key Takeaways

• The Model Hardware Standard is Anthropic's attempt to make one safety-aware interface the default for AI agents controlling physical machines, before foundation models are running robots at scale.
• Safety constraints — speed and angle limits, laser caps, emergency stops, human-approval gates — are written into the spec itself, and the models' physical reasoning is still weak enough that expert oversight is required, not optional.
• MCP's fast, near-universal adoption is the precedent Anthropic is counting on; whether rival labs and robotics firms take up MHS despite the friction it adds is the signal worth watching.

Frequently Asked Questions