Anthropic on Thursday opened a research preview of its Model Hardware Standard (MHS), a shared specification that lets AI agents discover, monitor and operate programmable physical equipment through a standardized software layer.
The company says MHS can connect devices including microscopes, liquid handlers, cameras, sensors and robotic arms, allowing agents to coordinate several machines at once. It is initially being tested by a limited group of scientific research labs and advanced manufacturers.
For security teams, the preview raises questions about how organizations will authenticate agents, restrict their permissions and monitor commands when AI systems can directly operate physical equipment.
The project began as a collaboration between Anthropic and HHMI Janelia Research Campus. Anthropic says MHS can reduce hardware integration work that typically takes weeks or months to hours or minutes, although the results come from Anthropic and organizations participating in the preview.
How MHS works
MHS uses a standardized driver between an AI agent and a physical device. Rather than requiring a separate integration for every piece of equipment, the driver exposes basic operations such as reading a temperature or changing a setting.
The system also gives agents structured information about a device’s capabilities and physical limits. Users can describe characteristics in natural language, including details that may not be obvious from software interfaces, and MHS can turn that information into a reference file for an AI agent.
Agents can then control equipment through Anthropic’s Model Context Protocol, command-line tools or code-based APIs. That allows an agent to coordinate multiple instruments, monitor results and adjust experimental parameters as conditions change.
In one example, Anthropic says Claude adjusted a laser, used a camera to observe the result and repeated the process until it understood how to align the beam. It then turned what it learned into a deterministic script that could perform the alignment without reasoning through every step again.
Why this matters
Anthropic is positioning MHS as a broader infrastructure layer rather than a Claude-only feature. The standard is model-agnostic and is intended to work with any device that has a programmable interface.
Organizations testing or developing support for MHS include Amazon Web Services, Automata, Danaher and Doosan Robotics. AWS plans to support MHS through Strands Robots, while Hugging Face is adding support to LeRobot and Raspberry Pi is enabling integrations across several products.
The potential applications range from automated drug-discovery experiments to robotics, manufacturing and quantum-computing systems.
Safety is the bigger test
Giving an AI agent control over physical equipment raises a different class of risks from letting it write code or manipulate information. A mistake can damage machinery, ruin an experiment or potentially create safety hazards. A compromised agent, account or driver could also expose connected equipment to unauthorized commands. Anthropic’s announcement discusses physical safeguards and model limitations but does not detail cybersecurity controls such as authentication, role-based permissions, network isolation, driver integrity or audit logging.
Anthropic acknowledges that its systems still have limitations. Claude’s understanding of the physical world can be incomplete, and researchers have had to intervene when the model misidentified physical problems as software issues. MHS also cannot currently support hardware without a programming interface.
The company is using the research preview to develop additional safety evaluations and deployment practices with participating partners and plans to open-source MHS after this testing phase.
The real opportunity may be interoperability
MHS could prove most valuable if it becomes a neutral layer between AI systems and the increasingly fragmented world of laboratory and industrial equipment.
Today, connecting machines from different vendors often requires specialist knowledge and custom software. A common interface could make that expertise less of a bottleneck and allow researchers to spend more time designing experiments rather than wiring them together.
But standardization alone does not make autonomous physical systems safe or secure. Before adopting MHS beyond controlled tests, organizations will need to determine whether it can enforce least-privilege access, record agent activity and fail safely when an agent or connected component behaves unexpectedly. Anthropic’s research preview may help establish those protections, but security teams should treat AI-to-hardware interfaces as critical control infrastructure rather than ordinary software integrations.