The former Google CEO warns the risks of artificial intelligence: “An example would be to learn how to kill someone”

Former Google general manager Eric Schmidt has launched a harsh warning on the dangers of artificial intelligence and its susceptibility to cyber attacks.

Schmidt, who drove Google between 2001 and 2011, drew attention to “Lthe bad ucors you can do have“, When asked if artificial intelligence can be more destructive than nuclear weapons, in an informal discussion at the” Sifted Summit “conference, according to CNBC.

Eric Schmidt said that there is “Evidence that models, whether they are open or closed, can be hated to lose their safety limits“, And in the process of training “I can learn how to kill someone.”

“There is evidence that you can take models, either closed or open, and hack them to eliminate the protection. During their training, they learn a lot of things. A negative example would be to learn how to kill someone,” Schmidt said.

“All major companies make it impossible for these models to answer such questions. Good decision. Everyone proceeds, and well. There is evidence that they can be reverse-English and many other similar examples,” he added.

The hazards have and vulnerability to attacks

AI systems are vulnerable to attacks, some methods being the injection of prompts and jailbreak. In a prompt injection attack, hackers hide malicious instructions in users or external data, such as web pages or documents, to deceive the AI ​​to do things that should not-such as to disclose private data or place harmful orders.

Jailbreaking, on the other hand, involves manipulating AI responses to ignore safety rules and generate restricted or dangerous content.

In 2023, a few months after the Chatgpt launch by Openai, users used a D trickIt’s “jailbreak” to bypass the safety instructions incorporated into the chatbot.

This included the creation of an alter-ego of Chatgpt called Dan (“Do Anything Now”), which involved the threat of death with death if he did not obey.

The alter-ego could provide information on how to commit illegal activities or list Adolf Hitler’s positive qualities.

Despite the dark warning, Schmidt has shown optimism to you in general and said that technology does not receive the attention it deserves.

“The GPT series, which culminated with the Chatgpt moment for us all, with 100 million users in two months, which is extraordinary, shows us the power of this technology.I, ”Schmidt said.