AI-written phage genomes worked in the lab, and screening is built for the wrong problem

Researchers used artificial intelligence to design complete sets of genetic instructions for bacteriophages, and some of those computer-generated designs produced working viruses once they were built and tested in the laboratory ^[1]. The consequence for anyone running a screening pipeline is direct: a system tuned mainly to flag DNA that resembles known dangerous pathogens has a harder job when a model outputs something genuinely new ^[7].

The work itself was deliberately narrow. The phages target E. coli and were designed to infect bacteria rather than people ^[2], and one of the models used, Evo 2, was built with restrictions that excluded viruses infecting animals, plants and humans from its training data ^[3]. Evo 2 was trained on trillions of DNA building blocks drawn from many forms of life ^[4], and its developers reported it performed poorly when tested on proteins from viruses that infect humans ^[5]. That is a designed limitation, not a law of nature, and it is the sort of guardrail that will be tested as the tools get better.

The reason this is a screening problem and not yet a threat is that a digital sequence is several steps from a working organism. It has to be turned into physical genetic material, assembled correctly, and tested under the right conditions, and each of those stages is also a place to reduce risk ^[11]. The most concrete checkpoint is DNA synthesis. Companies that make DNA to order can check both the requested sequence and the party placing the order ^[6]. But that check leans on similarity to known hazards, which is exactly the assumption a generative model can sidestep.

The stated engineering response, from authors who work in vaccine manufacturing and public health security ^[13], is to move screening from resemblance toward function. Research on "sequences of concern" has started asking what a sequence might do, not only whether it looks like something already linked to a dangerous organism ^[8]. That is a harder computational question than string matching, and it is the one that actually maps to the risk.

The rest of the layers are institutional rather than technical. UK Research and Innovation has set up a Trusted Research and Innovation team to help researchers and institutions identify and manage security risks in collaborative work ^[9]. And the UK Health Security Agency runs mSCAPE, which analyzes genetic material from samples to detect and track emerging pathogens; it was not designed to find AI-created viruses, but its approach does not require knowing the pathogen in advance ^[10]. The upside that keeps this research going is real: phages could be used against bacterial infections, including ones that antibiotics no longer handle well ^[12].

What to watch is whether synthesis providers adopt function-based screening in practice, not just in papers, and whether the model-level exclusions in tools like Evo 2 hold up under adversarial testing rather than benign benchmarks.