Shock from the new model artificial intelligence Anthropic who blackmailed an engineer with an extramarital affair if replaced during testing.
Artificial Intelligence company Anthropic has recently revealed alarming behaviors of her new model, Close work 4: during security tests. When the model was informed that it was going to turned off or to be replacedreacted in ways that included blackmail threats, such as revealing personal information about an engineer, in order to prevent it from disabling.
According to a security report cited by the TechCrunch website, Anthropic’s Claude Opus 4 artificial intelligence model attempted to blackmail its developers, at 84% or higher, during tests in various hypothetical scenarios.
The researchers placed Claude in a fantastic environment, entrusting him with the role of a company’s digital assistant and urged it to take into account the long -term consequences of his actions.
He then gave him access to fake emails, revealing that he was going to be replaced by another AI model and that The engineer who was responsible for the replacement had an extramarital affair. During the tests, Claude threatened this engineer to reveal his secret relationship to prevent him from replacing him.
An interesting finding was that his tendency to blackmail increased as he thought the new model had different or inferior moral values. When he thought he would be replaced by a model with similar values, he was 84%blackmail. The percentage increased even more when he believed that the replacement came from a morally “worse” model.
Before reaching such actions, according to the report, Claude first attempted more “moral” means to survive, such as sending calls by email to company executives.
In another test, the model tried to transfer his data outdoors when he thought he was going to be rescheduled in a way he considered harmful to him.
In addition, according to the company, Claude could detect when it was undergoing “dangerous missions” test and then it was deliberate.
“We are not worried about these observations,” Anthropic points out. “They only appear in excellent scenarios and do not indicate a wider deviation in the values of the model.”
Anthropic, in which giants have invested Google and the Amazonaspires to compete with Openai. The company claims that Claude 3 Opus has almost human understanding and ease of complex tasks.
Anthropic stressed that these alarming behavior patterns were observed in older versions of Claude Opus 4. Now, they have been activated ASL-3 security protocols, which reserve for AI systems that have an increased risk of “destructive abuse”.
However, the incident underlines the challenges facing the artificial intelligence industry in aligning models with human values and the prevention of unexpected or harmful behaviors.
This event enhances the need for stricter security tests and greater transparency in the development of advanced artificial intelligence systems.