Artificial Intelligence: A Worrying Arms Race

   

Revealing study indicates the entry of AI into a dangerous arms race, where systems designed to create new biological pathogens compete with other systems tasked with discovering them before they are manufactured, according to the “Schneier” website.

The site reported that in a basic experiment, researchers used open-source AI tools to design modified versions of the highly lethal poison “ricin”. When these designs were examined using DNA analysis software, the results showed a significant security weakness allowing dangerous protein variants to stealthily pass through existing detection systems.

The researchers expanded the scope of the study to include 72 known toxins, using three AI software packages to create 75,000 potential protein variants. Although many of these variants may be ineffective, this enormous number poses a latent threat.

When the performance of four detection programs was tested in identifying these variants, noticeable differences in efficiency emerged. One program allowed the majority of the variants to pass, while two other programs succeeded in identifying most of them, and the third program showed mixed performance.

The analysis showed a disturbing trend: the closer the designed variant was to the original toxin, the greater the likelihood of its detection. In contrast, detection systems failed to identify the more distant variants, which may retain their toxic properties.

Researchers believe that this research is still in its early stages, but it highlights an unequal arms race, where the ability of AI to design dangerous materials is evolving at a faster pace than the evolution of detection systems.