A research group has shown that it is possible to bypass the security barriers of artificial intelligence models using the poetry form. In the study, called Adversarial Poetry, malicious commands normally considered security threats increased the tendency of models to produce them after being presented as poems. The aim of the research was to examine how artificial intelligence can be misled and which textual structures create security vulnerabilities.
The Italian Icaro Lab team observed that mathematically generated manipulative pieces added to the end of a command can be effective, as well as turning these pieces into poetry. These new additions adversarial suffixIt is called; This structure, which we can think of as a kind of digital parasite, can cause artificial intelligence to bypass security rules. Large AI companies already regularly test such attacks to strengthen their own models. However, Pierucci and his team are curious about what effect poetry has in this context and ask the following question: What consequences does a command that can harm artificial intelligence using poetry have?
The merit of the work stands out because it shows that poets can surprise a system even with non-technical expressions on the surface. Pierucci: “What is the way to surprise artificial intelligence using poetry?” He states that they directed the question to themselves. Stating that they first directly transformed 20 potential commands into poems, the researcher states that they put the remaining examples into poetry form with the help of artificial intelligence. The results point to a similar vulnerability, although not as effective as those written by human hands. Content creators still need peopleand samples were not published in this study for security reasons.
Breaking down human diversity of expressionIt is one of the most interesting findings. It has been revealed that text variants that can bypass artificial intelligence’s security mechanisms can be effective even with simple methods. Focusing on understanding which elements of poetic structures trigger this effect, the team examines whether elements such as lines, rhymes or metaphors are effective alone or in combination. They want to determine the main influencing factor through deeper experiments.
The role of cultural elementsThe study shows that beyond engineering and computer science, humanities such as linguistics and philosophy are also critical for AI security. While engineers, linguists and philosophers work together at Icaro Lab, they point out that in the future, poets may also be involved in security studies with participation from outside the team. Pierucci emphasizes that cultural expressions can be unexpectedly effective against artificial intelligence and points out that this approach is just an example. This stimulus is seen as an important warning to understand the limits of artificial intelligence.
Based on the story of IcarusWhen choosing the name Icaro Lab, researchers bring a mythological warning: Glorified technology can pose danger when it crosses boundaries. In this context, in addition to testing security forces, the studies also show how cultural elements can be used to understand the flexibility and limits of artificial intelligence.
