DeepMind Puts a Watermark Inside AI-Designed Proteins

Google DeepMind has shown a way to hide a detectable signature inside proteins designed by AI, a step aimed at one of the harder problems in biosecurity. In results reported on September 30, the company said its SynthID watermarking system, already used to mark AI-generated images and text, could tag AI-designed protein sequences and later identify them with 100 percent accuracy at a false-positive rate of 0.1 percent, without changing how the proteins fold or bind.

The worry it addresses is concrete. AI tools can now design DNA sequences that behave like a dangerous pathogen without closely matching its genetic code, which means they can slip past the screening that DNA-synthesis companies use to catch orders for known threats. A reliable watermark would give those providers a second signal to check, one that does not depend on matching against a database of known-bad sequences.

Part of a larger push

The work grew out of a bioresilience program DeepMind launched with Isomorphic Labs on July 16, as the company described at the time. That effort splits into prevention, detection and response, and DeepMind says it has worked with more than 15 government agencies and biosecurity groups to shape the approach. The protein watermark sits in the prevention bucket.

It is still early. DeepMind has called the biology work exploratory rather than a finished product, and a watermark only helps if the tools that generate proteins actually apply it, which points to the harder question of whether rivals and open models will adopt the same scheme. A voluntary mark that only one lab uses does little to stop a determined bad actor.

Even so, the result is a rare piece of good news on the safety side of AI biology, a field that has mostly produced warnings. It also fits a wider turn toward putting AI to work in the lab, from Anthropic's new wet lab to Claude's recent work on enzyme systems. The tools that design biology and the tools that police it are, increasingly, the same tools.