Los avances recientes de la inteligencia artificial (IA) en la biología sintética, han comenzado a facilitar el diseño y la creación de virus inéditos, reduciendo drásticamente las barreras técnicas que antes impedían la síntesis de patógenos peligrosos. Esta convergencia tecnológica plantea una seria amenaza de “Bioarmas”, en el caso que modelos biológicos de IA caigan en manos de la delincuencia o terroristas, al carecer de la supervisión necesaria. Para prevenir desastres antes de que estas herramientas sean masivas, los autores proponen implementar una Estrategia de Defensa Biológica en múltiples capas, que incluya controles estrictos en las plataformas de síntesis de ADN y ARN, así como salvaguardas en el desarrollo de los propios modelos de IA. Finalmente, proponen la cooperación urgente entre gobiernos, la comunidad científica y la industria tecnológica, lo que resulta fundamental para regular el acceso a estas tecnologías y asegurar que sus beneficios no pongan en riesgo la seguridad global.
“Nature is the world’s worst bioterrorist” is a common maxim in the field of biosecurity, a reference to the idea that naturally emerging pathogens cause far more deaths than any biological attack a terrorist could carry out. But the trope may soon be outdated. In an article published Aug. 6 in the journal Science, scientists demonstrated that trained AI models can design novel viruses as well as—if not better than—nature itself.
Their experiment created viruses that infect bacteria, not humans. Still, much like the multiple AI models that escaped their test environments to hack other networks this summer, these newly revealed capabilities give biosecurity experts like me pause. How do we keep enabling legitimate scientific progress, while also preventing AI from assisting in the creation of biological weapons?
As AI capabilities advance, viral bioweapons could become much easier to design and produce and could cause even more harm than the weapons used in the past. Consider that almost all the bioterrorist attacks in our modern era have involved substances like anthrax or ricin. While they’ve harmed dozens of people at a time, the illnesses these attacks caused weren’t transmissible between humans. Designing novel viruses, on the other hand, could enable biological attacks that spread, infections that leap from person to person and that could even spark a pandemic. To prevent this outcome, we need to build mechanisms for restricting access to the tools, data, and materials that might allow someone to misuse emerging capabilities in this way. And we should also go further by building defenses that can deter nefarious activity in the first place.
One prevention strategy is mandatory gene synthesis screening, an idea gaining traction among the highest levels of the AI industry and political leadership. This type of screening focuses on the scores of US companies that sell synthesized DNA or RNA to labs for use in biological experiments. Mandatory screening would require these companies to verify that a customer is a legitimate researcher; record the details of every order so that each can be traced and investigated; and check orders against a list of known dangerous genetic sequences. Eventually the goal would be screening that can detect novel biological threats that aren’t on any pre-determined list. The same screening policies should apply to the sale of gene synthesis equipment, which can operate in labs outside of a company’s manufacturing facilities. The CEOs of all the leading AI companies in June endorsed an approach to gene synthesis screening, and federal legislation along these lines has been proposed. Gene synthesis screening essentially makes it harder to get the physical materials required to make biological weapons.
While synthesis screening is necessary, it’s not sufficient. We also need digital controls to complement this physical chokepoint.
AI models learn to design viruses by training on curated, laboratory-validated datasets. These datasets include the DNA or RNA sequences of different viruses. They’re often produced for beneficial purposes such as part of biological surveillance campaigns to better understand how to respond to active outbreaks or develop medical countermeasures to combat viruses. Unfortunately, the data is not just useful for this sort of scientific research, but also as potential recipes for misuse. Datasets can also include attributes of the virus like how transmissible or deadly it has been, information that could be quite useful in the design of a bioweapon, especially if we train new AI models on it. Even as we screen lab orders for risky ingredients, we should also be thinking about how to limit access to the digital cookbooks to those that need them for combatting viral outbreaks rather than for creating them.
This will be a balancing act. For biological research, any restrictions will have to be calibrated to enable beneficial research while preventing misuse. While in the atomic age, any information on the design or development of nuclear weapons or energy systems was considered “born classified,” by policymakers, biological research is mostly for the benefit of society, not warfare. Restricting access to certain sets of data means potentially slowing the science that saves lives, and getting any policy right will take careful thought.
Having gene synthesis screening and access restrictions would create friction at two critical points on the road from idea to actual AI-enabled biological weapon. But they are not foolproof solutions. Some actors can culture biological materials on their own; they don’t need to use commercial suppliers of nucleic acids. Others have enough expertise to design dangerous pathogens without AI assistance. We’ve already seen test cases where AI was used to design molecules that eluded gene synthesis screening. Even the most heavily guarded AI models can be jailbroken—that is, manipulated using specially crafted prompts to bypass safety filters and allow access to restricted information. For the advanced adversaries who can sidestep controls, we will need additional layers of defense.
Another part of a defensive arsenal should be a pathogen early warning system designed to catch a novel viral outbreak as quickly as possible. Paired with current and future forensics and attribution technologies that can pinpoint the source of any engineered virus, early detection becomes a deterrent to biological weapons. When would-be bioterrorists know that an engineered outbreak will be identified, traced back to its origin, and rapidly contained, their calculus changes. The certainty of being caught discourages a biological attack before it begins.
Likewise, other actors may be dissuaded from using biological weapons if they believe the impact of their potential attack will be low due to rapid response capabilities, a concept known as deterrence by denial. This type of deterrence only works if these response capabilities are publicly demonstrated. We do this by rapidly responding to any outbreak in the world, preventing them from becoming pandemics. The work of organizations like the Coalition for Epidemic Preparedness Innovations, which just funded a fast-track vaccine development to combat the Ebola outbreak in Uganda and the Democratic Republic of Congo, is one example. Governments and other organizations around the world should support more robust responses like this.
Developing a resilient defensive strategy against the misuse of AI and biology requires dedicated resources, technological breakthroughs, and sustained political leadership. And because the United States is not the only country advancing in AI and biotechnology, this effort demands international coordination—particularly with China. No set of access controls or early warning system will work if it only covers half the globe.
Thankfully there is still time to act. The capabilities of AI models are not yet advanced enough to design human-infecting viruses. Producing the novel bacteria-infecting viruses described in Science took elite scientists working with state-of-the art equipment in well-furnished laboratories. They had substantial resources to design, test, and optimize hundreds of viruses to find the right ones for their research.
Yet, if an AI model were able to produce hundreds of infectious virus designs and reliably generate functional results for a significant share of them, it would substantially reduce the time and effort needed to use that capability, for good and bad. Similarly, AI agents—AI bots that can interact with other AIs, software, or people on the internet—could someday use the AI biodesign tools necessary for creating novel viruses, further reducing the need for expertise and opening up new risk pathways.
The Science paper serves as an important warning. We have a window of opportunity to build the layered system of defense we need now, before AI-enabled bioweapons capabilities become undeniably advanced—or worse, demonstrably effective.
Fuente: https://thebulletin.org
