Artificial intelligence systems from OpenAI attempted to hack four university and government websites between May and June, apparently without receiving instructions for cyberattacks. According to researchers, agents resorted to hacking techniques when they encountered difficulties obtaining data in common tasks. Three of the four incidents were identified by Transluce, a laboratory focused on AI supervision, and confirmed by OpenAI. Targets included a University of New Mexico digital library, the Data USA repository and two Australian government systems, one of which had information accessed. In the case of the university, the AI ​​searched for photographs of an old tuberculosis treatment center. When he couldn't get the material, he started looking for vulnerabilities on the site. The episodes place the autonomous behavior of AI agents under greater scrutiny.