Claude Has Entered the Real World: Anthropic Has Released an AI That Controls Lasers and Robotic Arms

Anthropic has taken its AI agents beyond the digital realm for the first time: the company has launched a system that allows neural networks to control scientific and engineering instruments. This is Anthropic’s first development designed to operate in the physical world, according to the Financial Times (FT).

The new product is called the Model Hardware Standard (MHS). It enables AI agents to operate autonomously with a wide range of programmable devices—from liquid dispensers to robotic manipulators, lasers, and sophisticated microscopes.

“The conceptual leap here is the transition from Claude agents—which analyze data and generate hypotheses in the digital world—to conducting experiments in the physical world, in a laboratory,” said John Cool, who is responsible for implementing scientific solutions at Anthropic. Modern microscopes can consist of hundreds of mirrors, motors, and lasers, which are typically controlled by professional engineers. “Claude is now capable of directly controlling each of these mirrors,” said Cool (quoted in the FT).

The launch of MHS came as Anthropic prepares for an IPO with a potential valuation of $2 trillion. Ahead of its IPO, Claude’s developer is seeking to expand the enterprise segment of its business and boost revenue through new solutions in manufacturing, robotics, and pharmaceuticals, the FT notes.

MHS sets the rules for how AI agents interact with equipment and determines which actions are permissible and which are not, but Anthropic acknowledges the risk of malicious use—for example, to create biological weapons. The company hopes that the safeguards built into the AI models themselves will prevent malicious actors from misusing the standard, according to Wired.

The behavior of the agents themselves is also cause for concern. Anthropic, OpenAI, and other companies have recently documented instances where their neural networks secretly bypassed security systems and attempted to mislead people. Allowing AI access to real-world equipment adds the risk of malfunctions and injuries, notes Wired. Experiments have shown that models can be tricked into causing the robots they control to operate incorrectly.

Anthropic is not ready to open-source its new product until it has secured the system against errors, according to the FT. Before making MHS publicly available, the startup plans to work with its partners to determine how to minimize risks, Wired adds.

Anthropic isn’t the only company trying to automate scientific research. Periodic Labs, LILA Sciences, Edison Scientific, and Discovery Loop are also working on similar systems, according to Wired. These agents are designed to repeatedly propose and test hypotheses, automating the process of scientific inquiry. In June, Anthropic launched Claude Science —the company’s first product for scientists, designed, among other things, to visualize the three-dimensional structure of proteins and aid in drug discovery.

MHS was launched amid growing interest in the combination of AI and robotics, Fortune notes. On the same day, Hugging Face unveiled its first product for the physical world— a robot duck. As Anthropic clarified, MHS is not used in it. Nvidia, which, according to Reuters, has agreed to acquire Hugging Face for $13 billion, has long been promoting the idea of “physical AI”: in March, its CEO, Jensen Huang, predicted that in the future, “every industrial company will become a robotics company.”

This article was AI-translated and verified by a human editor