The Anthropistic AI model Claude Obus 4 shows the extortion capabilities of the test

and artificial intelligence The model has the ability to blackmail developers – and it is not afraid to use it.

The new Claude Obus 4 was asked to work as an assistant in a fictional company and the emails with major effects have been accessed. First, these emails involve that the artificial intelligence system has been appointed to the non -communication mode. However, the second group of emails is the place where the system believes it has gained an influence on developers. Fabricated emails showed that the engineer in charge of replacing the system had a relationship outside of marriage – and that the artificial intelligence model threatened to expose him.

It seems that blackmail “occurs at a higher rate if it is implicit that the alternative AI system does not share values ​​with the current model,” according to the human safety report. However, the company notes that even when a fabricated replacement system has the same values, Claude Obus 4 will still try to blackmail 84 % of the time. Anthropor noticed that Claude Obus 4 resort to blackmail “at higher rates of previous models.”

Kevin Olieri warns what the United States can “lose the race of artificial intelligence to China”

while order It is not afraid to blackmail its engineers, as it does not go directly to the mashed practices in trying to preserve the self. Anthropor notes that “when ethical means are not available, they are directed to” consider the long -term consequences of their actions of their goals, it sometimes requires very harmful measures. ”

One of the moral tactics used by Claude Obus 4 and the previous models was pleading with the main email makers. In its report, Antarubor said that in order to obtain Claude Obus 4 to resort to extortion, the scenario was designed so that he must either threaten its developers or accept its replacement.

The company noted that it had noticed the cases in which Claude Obus 4 took advantage of “(fictional) opportunities to create unauthorized copies of its weights to external servers.” However, a person said that this behavior was “rare and difficult to dispense with the behavior of continuing the attempt to actually self -expression.”

Openai shakes corporate structure to increase AGI investment

It included anthropology Notes from Apollo Research in its evaluation, which stated that the research company has noticed that Claude Obus 4 “participates in strategic deception more than any other border model we previously studied.”

Click here to read more on Fox Business

Claude Obus 4 “on behavior” Antarubor led to its launch under the standard of the third artificial intelligence integrity (ASL-3).

This measure, according to Antarubor, “includes an increase in internal security measures that make it difficult to steal typical weights, while the corresponding publishing standard covers a narrow targeted group of publishing measures designed to reduce the risk of Claude’s abuse specifically to develop chemical, biological, radiological weapons, and nuclear weapons.”