Eric Schmidt, the former CEO of Google, has sent out a clear warning about the risks associated with artificial intelligence and how easily it may be compromised.
During a fireside talk at the Sifted Summit, Schmidt, the former CEO of Google from 2001 to 2011, cautioned about “the bad stuff that AI can do” when asked if AI is more harmful than nuclear weapons.
Could there be an issue with AI proliferation? "Yes," Schmidt stated on Wednesday. One of the hazards associated with AI proliferation is that it could be misused and repurposed by malicious actors.
There is proof that models, whether closed or open, may be hacked to remove their guardrails. They thus pick up a lot of knowledge during their training. "Learning how to kill someone would be a bad example," Schmidt stated.
"Those models are unable to respond to that query because of the actions of all the major corporations. Well done. This is something that everyone does. They perform it effectively and with the proper motivations. They can be reverse-engineered, according to evidence, and there are numerous other instances of this kind.
Attacks on AI systems can take several forms, including as jailbreaking and prompt injections. In order to fool the AI into doing things it shouldn't, like sharing private information or executing dangerous orders, hackers conceal malicious instructions in user inputs or external data, such as web pages or documents, in a prompt injection attack.
Contrarily, jailbreaking entails tampering with the AI's reactions to cause it to disregard its safety guidelines and generate content that is banned or hazardous.
A few months after OpenAI's ChatGPT was made available in 2023, users used a "jailbreak" technique to get around the chatbot's built-in safety guidelines.
This involved developing a ChatGPT persona called DAN, which stands for "Do Anything Now," and threatening to kill the chatbot if it didn't comply. The alter ego can enumerate Adolf Hitler's virtues or offer advice on how to engage in unlawful activity.
According to Schmidt, there is currently no effective "non-proliferation regime" to assist reduce the risks associated with AI.
