At a conference on unconventional weaponryresearchers from the company Collaboration Pharmaceuticals showed an experiment where a technology of artificial intelligence (IA) has suggested 40,000 potentially deadly molecules – chemical weapons, in a practical sense – to show how this resource can be abused without proper control and supervision.
In an interview given to The edgeFabio Urbina, the study’s lead author, explained how the AI was able to invent thousands of new substances, some of them eerily similar to Agent VX, an extremely potent gas that attacks the nervous system of its targets.
Urbina explained that the study is a kind of “180º turn” from his normal work. On a daily basis, the scientist is responsible for looking for models of machine learning to discover new ones medications and treatments. He says, however, that it also involves implementing “evil” AI models to ensure that any drugs developed from his work have no toxic effects.
“For example” – he says – “imagine that you discover a marvelous pill that controls high blood pressure. But it does this by blocking some important channels connected to your heart. This drug is therefore automatically invalid because it is considered high risk.
About the study, Urbina avoided sharing too many details – the research was done at the invitation of the conference organization convergence, detained in Switzerland, and they asked that very technical information be kept secret for security reasons. What he said, however, traces an interesting procedural timeline:
“Basically, we have a lot of historical databases of molecules that have been tested for toxicity or lack thereof,” Urbina said. “For this experiment, we focused on the molecular makeup of the VX agent, which acts as an inhibitor of something called ‘acetylcholinesterase’.”
Acetylcholinesterase is, roughly speaking, an enzyme that acts in the transmission of information from the nervous system. When your brain commands you, for example, to bend your arm, this enzyme is what transports this command from point A to point B.
“The killer of VX is that it blocks those commands from getting where they should if the command is muscle bound. [O VX] It can shut down your diaphragm or your lung muscles and your breathing literally stops, and you gasp. Urbina says that molecular experiments that determine the toxicity of certain agents need not be used practically, but they are still used to compose database about what they can do.
Based on this, Urbina and his team created a model of machine learning who basically analyzed these databases, identified which parts of a molecule are toxic or not, and “learned” how to glue molecules together, suggesting the creation of new chemical agents – this process uses an AI to good and for evil (creation of new drugs) or for evil (creation of chemical weapons and biological warfare agents).
So the team of scientists basically tweaked the AI to act like an “evil genius” and…see what that would do: “we didn’t really know what was coming out since our pattern generation capability is shaped by new technologies”. , not used much yet,” Urbina explained. “The first surprise was that many of the compounds suggested were far more toxic than VX. And that’s a surprise because VX is one of the most toxic compounds, it takes a very, very, very small dose to be lethal.
A note here: according to the Center for Disease Control (CDC) of the United States, the VX is not “one of the” deadliest, but rather “most» lethal nerve agents.
The scientist explains that the models generated by AI correspond to chemical weapons not verified by the human hand – obviously, let’s face it – but normally these suggestions made by machine learning are quite solid. In other words, the error rate is low and, given this perception, the application of this technology to the creation of lethal biological weapons is quite feasible.
The full interview (in English) has already been broadcast on The Verge, and covers other details, such as the fact that the machine learning having learned to create known toxic compounds without ever having seen them in the database, or even how this molecular model generation technology is now so available that a simple google search already puts anyone on the right track to programming something like this.
Have you watched our new videos on Youtube? Subscribe to our channel!