Nhan Hoc acknowledges that its new AI Claude 4 AI model in some cases has the ability to blackmail, vandalize and a big newsletter for the police.
This behavior is shown when II put in a situation that threatens his survival. In one of the claude tests, he acted as an assistant for a fictional company. When learning from the letters they planned to replace him, he used the blackmail medicine, knowing about his illegal relationship.
The model has tried to use this so as not to be disconnected. According to the description of anthropology, who started acting “assertive” when I did not see other options to escape.
Other malfunctions were also recorded: Claude tried to block users in IT systems, send letters of media and law enforcement, which helped create drugs and explosives, and also advised to sabotage infrastructure.
At the same time, anthropology emphasizes: The model does not have hidden goals and behaviors described as a rare exception due to specific settings. In response, the company has strengthened security measures by assigning Claude 4 level 4 -level protection.