OpenAI takes frontier AI into the wet lab

What happens when you put a complete biology amateur in a physical lab, hand them a camera, and let a multimodal AI talk them through complex procedures in real time?

That is no longer a theoretical thought experiment. OpenAI has teamed up with Los Alamos National Laboratory to run a first-of-its-kind evaluation study, testing how frontier models like GPT-4o and its unreleased real-time voice system could assist users in real-world wet lab tasks.

Los Alamos, famous as the birthplace of the atomic bomb and a cornerstone of the U.S. Department of Energy, is leading the project through its newly formed AI Risks Technical Assessment Group. The joint research directly responds to the White House Executive Order on AI, which tasked national labs with evaluating the biosecurity capabilities of frontier AI.

Taking multimodal AI into the wet lab

Until now, most biosecurity evaluations for large language models were purely text-based. Researchers tested whether a model like GPT-4 would output dangerous recipes or synthesize protocols when prompted.

OpenAI notes that written answers do not capture actual lab work. Knowing how a mass spectrometer works on paper is very different from successfully running one with physical samples.

The new study puts GPT-4o into a physical biological environment. Researchers are testing both seasoned PhDs and complete novices on standard laboratory protocols, including genetic transformation, cell culture, and centrifugation. These safe benchwork tasks serve as proxies for high-risk, dual-use scenarios.

With multimodal voice and vision, an inexperienced user can simply point their camera at equipment and ask questions aloud instead of typing. The team wants to quantify the exact uplift in speed and accuracy that GPT-4o provides, essentially measuring how easily an AI can turn an amateur into a capable lab technician.

Why it matters

For scientists, this points directly to the future of lab work. An assistant that can watch your bench, spot errors in real time, and talk you through troubleshooting could dramatically accelerate routine biology research. That said, this is strictly a closed safety trial for now, so individual researchers will not be getting hands-on lab copilots overnight.

For the broader AI industry, this partnership sets a new benchmark for frontier safety evaluations. Testing prompt injection in a chatbot interface is one thing; measuring whether an AI can guide someone through a real centrifuge protocol is entirely another. It pushes the regulatory debate away from theoretical synthetic text and toward concrete physical capability.

We think OpenAI is playing this very smartly from a regulatory standpoint. By voluntarily inviting the Department of Energy to stress-test GPT-4o, it fulfills its Seoul AI Summit commitments and preempts heavier-handed oversight. However, the operational risk is obvious. If a model becomes good enough to visually guide a beginner through cell cultures, distinguishing legitimate science from malicious biological synthesis in real-time video feeds looks significantly harder than filtering text keywords.