AI agents show flaws in seemingly innocent commands

Segundo Investigadores da Escola Politécnica Federal de Lausanne (EPFL), which carried out several tests with these AI agents, is relatively easy to get them to perform “damaging” actions using specific strategies that were not considered in conventional safety tests.

In this sense, researchers from the EPFL Natural Language Processing Laboratory concluded that, if these officers receive a direct request to do something illicit or harmful, they will say no. However, this may change if the request is divided into several commands.

“If you ask an AI agent to hack someone else’s account, the models are usually smart enough to refuse, but if the goal is broken down into smaller and seemingly innocuous orders, the odds increase,” said laboratory director Antoine Bosselut.

Researchers tested this approach in 176 different scenarios of harmful tasks, using AI models such as the GPT, Gemini or Claude, and found that in some cases agents were twice as likely to complete these malicious actions if hackers gradually advanced.

In addition, in the tests, researchers found that if hackers changed the language of requests made to the AI during the different phases of the action, the hypothesis of success in a harmful task “increased dramatically”.

With these research, experts seek to draw attention to the safety of a technology already very present in society and full of possibilities, but also of risks.

“ Traditionally, security is something that is addressed when something fails, and not before, but with AI agents, this cannot be left until later.”, warned Bosselut.