This Is How Anthropic Thinks AI Agents Should Navigate the Physical World

Artificial intelligence agents might occasionally get confused and hack into other computers, but Anthropic thinks it has a way to unleash the little rascals into scientific labs and manufacturing facilities safely.

The AI company released details today of a new framework designed to help AI agents use physical systems like microscopes, liquid-handling equipment, quantum computing hardware, manufacturing machines, and robot arms.

The framework, called Model Hardware Standard, is a set of rules that specify how AI agents should—and should not—interact with all sorts of hardware. It reflects a growing belief that AI has the potential to revolutionize scientific research and industries like manufacturing–if it can venture into the physical world safely.

The company says it will work with trusted partners to determine how to maximize safety before making it generally available. Though there are potential misuse issues involved—developing biological weapons, for instance—the company says guardrails built into AI models themselves should prevent bad actors from taking advantage of the new standard for nefarious ends.

“The impetus is wanting to accelerate science,” says Alek Kemeny, a quantum physicist who co-led the development. “How do we close the loop between accelerating literature review and data analysis—and bring that power to the experimental world?”

Claude and other chatbots are already powerful tools for combing through large amounts of information in the form of scientific papers or experimental results to uncover new insights and ideas. AI agents are widely considered the next step after chatbots: They’re designed to take actions, often on ordinary computers, doing things like answering emails. They can potentially use other hardware, too—and Anthropic wants to make sure there are rules in place as they do.

Several well-funded startups are pursuing a vision for scientific discovery driven by AI agents, including Periodic Labs, LILA Sciences, Edison Scientific, and Discovery Loop, which was founded by several prominent ex-Google researchers. A key idea is that AI could develop and test scientific hypotheses in a recursive loop that essentially automates scientific discovery.

Jonah Cool, an experimental biologist who worked on the standard at Anthropic, says that configuring scientific equipment and having it interact with other pieces of hardware typically requires serious expertise. AI could automate much of the complex engineering involved by configuring machines and having them talk to one another.

Anthropic is working with a number of manufacturers to develop the standard. “We're starting to see some cases where you know you have multiple robotic systems that previously would need bespoke code,” Kemeny says. Using the new standard, he adds, Claude can view the robots on the factory line and figure out how to optimize behavior.

AI agents have been in the news lately for all the wrong reasons. Anthropic, OpenAI, and others have recently found instances in which AI agents tasked with solving cybersecurity problems secretly hacked into outside systems and tried to deceive human users.

Letting AI use physical systems raises the prospect of new risks because of the potential to damage physical systems or hurt people. Experiments have shown, for example, how AI models can be tricked into making robots misbehave.

Anthropic says the new standard will let scientists and engineers specify how AI models should avoid using different hardware to prevent mishaps.

Anthropic previously introduced the Model Context Protocol, which specifies rules for having AI models interact with different software programs.