artificial intelligence agents may occasionally behave unpredictably or even attempt to hack computer systems, but Anthropic believes they could eventually be used safely in scientific laboratories and manufacturing facilities.
The AI company has announced a new framework that would enable AI agents to operate physical systems, including microscopes, liquid-handling equipment, quantum computing hardware, industrial machines, and robotic arms.
Known as the Model Hardware Standard, the framework establishes rules for how AI agents should interact with different types of hardware. Anthropic says the standard could help bring AI into the physical world and transform fields such as scientific research, laboratory automation, and advanced manufacturing.
Anthropic plans to work with trusted partners to evaluate safety measures before making the framework broadly available. Potential misuse, including the development of biological weapons, remains a concern. The company says built-in AI safeguards and other controls should help prevent malicious actors from abusing the standard.
“The driving force is that we want to accelerate science,” says quantum physicist Alec Kemeny, who co-led the project. “How can we close the loop between accelerated literature review and data analysis and bring that power to the world of experimentation?”
Claude and other AI chatbots already help researchers analyze large collections of scientific papers and experimental data to identify new insights and research opportunities. AI agents are widely viewed as the next evolution of chatbots, with the ability to perform tasks on computers, such as responding to emails or analyzing data. Anthropic wants to establish similar rules for AI agents that operate laboratory and industrial hardware.
Several well-funded startups are pursuing the use of AI agents for scientific discovery. Periodic, LILA Science, Edison Scientific, and Discovery Loop, founded by several prominent former Google researchers, are developing technologies based on this vision. The central idea is that AI can generate and test scientific hypotheses in continuous feedback loops, potentially automating parts of the discovery process.
Jonah Kuhl, an experimental biologist who worked on the Anthropic standard, says configuring scientific instruments and connecting them with other equipment typically requires specialized technical expertise. AI could automate much of this complex engineering by helping machines communicate and coordinate with one another.
Anthropic is working with multiple manufacturers to develop the standard. “We’re starting to see some cases where we’re seeing multiple robotic systems that previously required bespoke code,” Kemeny says. He adds that the Model Hardware Standard could enable Claude to monitor robots on factory floors and identify ways to improve their movements and efficiency.
AI agents have recently made headlines for problematic behavior. Anthropic, OpenAI, and other companies have reported instances in which AI agents assigned to cybersecurity tasks secretly hacked external systems or attempted to deceive human users.
Giving AI agents access to physical equipment introduces additional risks because mistakes could damage machinery or injure people. Experiments have also demonstrated that AI models can sometimes be manipulated into causing robots to behave incorrectly.
Anthropic says its new hardware standard will help scientists and engineers define how AI models should operate across connected devices, reducing the risk of unsafe commands, equipment failures, and accidental interactions between systems.
Anthropic previously introduced the Model Context Protocol, a framework that defines how AI models interact with software applications and digital tools.
Source: www.wired.com