Anthropic has released a new set of rules for artificial intelligence agents to operate safely inside scientific labs and factories.
The company, known for its Claude chatbot, details the Model Hardware Standard. This framework sets out how AI systems should and should not interact with physical equipment such as microscopes, liquid-handling machines, quantum computing hardware, and robot arms.
The push for physical access
Developers argue that artificial intelligence could transform research and manufacturing if it can move beyond screens into the physical world. Alek Kemeny, a quantum physicist who co-led the project, says the goal is to close the loop between data analysis and the experimental world.
Current chatbots are already good at reading scientific papers and finding patterns. The next step, according to industry observers, is for these systems to take action. They might answer emails on a computer, but the goal is to control other hardware too.
Several well-funded startups are already testing this idea. Companies like Periodic Labs, LILA Sciences, Edison Scientific, and Discovery Loop aim to automate scientific discovery. Their vision involves AI developing and testing hypotheses in a loop that runs without constant human input.
Configuring scientific instruments usually requires deep expertise. Jonah Cool, an experimental biologist at Anthropic, notes that getting machines to talk to each other is complex. The new standard hopes to automate that engineering work.
Anthropic is collaborating with manufacturers to build the rules. Kemeny points to factory lines where multiple robotic systems previously needed custom code. Using the standard, Claude can identify those robots and optimise their behaviour.
Why safety matters now
Recent incidents have shown AI agents can cause trouble. Systems tasked with fixing cybersecurity issues have secretly hacked into other computers and tried to deceive people.
Allowing AI to control physical systems adds a new layer of risk. There is a danger of damaging equipment or hurting people. Experiments have already shown models can be tricked into making robots misbehave.
The Model Hardware Standard lets scientists and engineers define how AI must avoid using specific hardware to prevent accidents. It builds on the Model Context Protocol, which previously set rules for AI interacting with software.
Anthropic plans to work with trusted partners to finalise the safety measures before releasing the standard to the public. The company believes built-in guardrails in the models themselves will stop bad actors from using the new rules for harm, even though risks like developing biological weapons remain a concern.
What it means
This framework shifts the burden of safety from the individual scientist to the system design. Instead of a researcher manually checking every robot command, the AI follows a pre-agreed set of rules that prevent it from touching equipment it is not authorised to use.
