On Thursday, a group of researchers from Microsoft announced the identification and potential mitigation of what they are calling a biological zero-day. This term refers to an unrecognized vulnerability in a system designed to guard against biological threats. The risk lies in a system that monitors purchases of DNA sequences to identify orders related to toxins or hazardous viruses. The researchers assert that this system is now increasingly susceptible to overlooking a new kind of danger: toxins engineered by artificial intelligence.
Determining the scale of this threat requires a deeper understanding of current biosurveillance initiatives and the functions of AI-generated proteins.
Identifying Potential Dangers
Biological threats manifest in various ways. They can include pathogens like viruses and bacteria, protein-based toxins such as ricin—which was famously mailed to the White House in 2003—and chemical toxins produced through enzymatic processes, like those found in red tide. All these threats originate through a fundamental biological mechanism where DNA is transcribed into RNA, subsequently leading to protein formation.
For many years, initiating this biological process has become as straightforward as ordering the desired DNA sequence online from various companies, which will fabricate and ship the requested sequences. In light of the potential risks associated with such ease of access, both government agencies and the private sector have collaborated to implement a screening phase for every order. This involves scanning the DNA sequence for its potential to encode components of recognized protein toxins or viruses. Sequences flagged during this process are then escalated for human evaluation to determine if they pose a genuine threat or if the individuals placing the orders are dangerous.
The list of proteins of concern, as well as the sophistication of the screening methods, has been constantly refined over the years to keep pace with research advancements. Early screenings focused primarily on DNA sequence similarities. However, since multiple DNA sequences can produce the same protein, the screening algorithms have evolved to account for various DNA variants that can represent the same risk.
