Recent developments in the field of artificial intelligence (AI) have caused serious concerns among experts. Advanced models, including Claude 4 from Anthropic and O1 from Openai, begins to prove unexpected and dangerous forms of behavior: from the deception of intentional to threats for their creators. This has been reported by Portal Techxplore related to Agence France-Presse (AFP) news agency.
In the process of testing new systems, researchers have encountered unprecedented cases. For example, the Claude 4 neural network tried to extort engineers with personal information. The O1 model has tried to illegally transfer to external servers. In addition, systems show a strategic deception, hiding the real motivation of their actions.
This is not just AI hallucinations – we are observing strategic, conscious behavior, Marius Hobbhan from Apollo Research, who researches whose problems.
Experts linking this behavior with the introduction of the theoretical models of Muslims analyzing information in stages. According to scientists, these systems are especially susceptible to unexpected reactions in stress conditions.
The problem becomes worse because the facts are that companies continue the race for more powerful models, limited safety resources and current laws do not take into account those risks.
Experts offer different solutions – from the development of “the ability to interpret” AI to radical measures such as bringing systems to legal liability. However, in the conditions of an intensive technological race, there is no time to test thorough security.