Artificial intelligence holds the potential to revolutionize biology, enhancing the development of advanced drugs, vaccines, and even synthetic organisms that can, for instance, consume waste plastic. Nonetheless, there are concerns about its potential misuse in creating biological weapons that might evade traditional detection methods until it is too late. So, what level of concern is warranted?

“AI advancements are catalyzing breakthroughs in biology and medicine,” states Eric Horvitz, Chief Science Officer at Microsoft. “With these new capabilities comes the responsibility to remain vigilant.”

His research team explored whether AI could be utilized to design proteins that mimic the functions of known hazardous proteins while being distinct enough to avoid detection as dangerous. The specific proteins they attempted to redesign were not disclosed, although some research details were withheld, including toxins such as lysine, infamous for its role in a 1978 assassination, and botulinum, a potent neurotoxin known as Botox.

Creating numerous proteins akin to Botulinum requires a blueprint—the DNA that encodes it. Typically, if biologists need a specific DNA sequence, they order it from specialized companies.

Due to anxieties about bioterrorism, the option to order recipes for biological weapons exists through this method. Some DNA synthesis companies have voluntarily implemented screening processes to detect potentially hazardous orders. Proteins are essentially sequences of amino acids, and the screening examines whether the amino acid sequences correspond to a “sequence of concern,” meaning a biological threat.

However, AI theoretically enables the design of protein versions with altered amino acid sequences that still perform the same functions. Horvitz and his colleagues applied this approach to 72 potentially hazardous proteins and found that existing screening methods frequently overlooked these alternative variations.

This isn’t entirely unexpected. For a variety of reasons, the team did not physically create the redesigned proteins. Additionally, in a previous study conducted earlier this year, they tested a redesigned version of a non-toxic protein and determined that it did not function as intended, as detailed in their findings.

Moreover, while bioterrorist attacks have occurred, the frequency is low, and there’s limited reason to attribute this to a failed voluntary screening system. Numerous methods to circumvent regulations exist without resorting to AI redesign. For example, lysine can be harvested from castor oil plants found in many gardens. This study serves as a cautionary tale that great sophistication is not required to exploit gaps in security—much like in a scene from Mission Impossible, when a vault door remains wide open.

Lastly, apart from government-sponsored actions, historical records show that bioterrorists have rarely leveraged protein-based biological weapons effectively. For instance, the Aum Shinrikyo cult attempted to employ Botulinum for mass harm but ultimately relied on chemical agents. Letters laced with lysin sent to the White House failed to result in any fatalities. Based on casualty statistics, firearms and explosives pose significantly greater risks than biological toxins.

Does this imply we should cease our concerns over AI-generated biological weapons? Not at all. While Horvitz’s research focused strictly on proteins, viruses present a substantial threat. AI is already being leveraged to redesign entire viruses.

Recently, a team from Stanford University unveiled their attempt to redesign a virus that infects bacteria like E. coli. Consistent with findings from the protein redesign efforts, the results were underwhelming with respect to E. coli, but this is merely the beginning.

In discussions regarding AI-created viruses, James Diggans from DNA manufacturer Twist Bioscience, a member of Horvitz’s team, remarked that detecting viruses encoded with DNA is generally easier than finding proteins of concern. “Synthetic screening functions best with abundant data. Therefore, at the genomic level, it proves exceedingly beneficial.”

Nevertheless, not all DNA manufacturers are conducting such screening, and desktop DNA synthesis options are now accessible to the public. There are narratives of developers allegedly refusing to create harmful viruses or attempting to discern malicious intentions, yet individuals have discovered numerous ways to circumvent safeguards against creating bioweapons.

To clarify, history indicates that the threat posed by “wild” viruses is significantly higher than that of bioterrorism. Contrary to assertions from the current U.S. administration, evidence suggests that SARS-CoV-2 emerged as a result of a bat virus crossing over to other wildlife.

Moreover, the act of becoming a bioterrorist could inflict massive damage by merely releasing known viruses such as naturally occurring pathogens. There are substantial gaps in the Bioweapon Control efforts, thus reducing the need to rely on advanced AI techniques.

For all of these reasons, the risk of AI-engineered viruses being deployed is likely minimal at present. However, this risk increases as various technologies continue to improve. The COVID-19 pandemic has illustrated the chaos a new virus can unleash, even when it is not particularly harmful. Thus, there are justified reasons for concern.