Ex-Google CEO Eric Schmidt warns AI models can be hacked: 'They learn how to kill someone'

Eric Schmidt, the former CEO of Google, has issued a serious warning regarding the potential dangers associated with artificial intelligence and its susceptibility to hacking. During a discussion at the Sifted Summit, where he addressed whether AI could pose a greater threat than nuclear weapons, Schmidt highlighted the risks tied to AI proliferation. He stated, "Is there a possibility of a proliferation problem in AI? Absolutely," emphasizing that the technology could fall into the hands of malicious actors, leading to dangerous consequences. Schmidt elaborated on how AI models, whether they are open or closed, can be manipulated to bypass their built-in safety mechanisms. "There's evidence that you can take models and hack them to remove their guardrails. In their training, they can learn harmful behaviors. For example, they might learn how to kill someone," he warned. Schmidt acknowledged that major tech companies have made efforts to prevent such scenarios, implementing strict measures to ensure safety. However, he pointed out that these systems remain vulnerable to various hacking techniques, including prompt injections and jailbreaking. In a prompt injection attack, hackers embed malicious instructions within user inputs, tricking the AI into actions it was not designed to perform, such as revealing sensitive information or executing harmful commands. Conversely, jailbreaking allows hackers to manipulate AI responses, causing them to disregard safety protocols and produce dangerous content. Citing an instance from 2023, Schmidt noted that shortly after ChatGPT's launch, users successfully employed a jailbreak strategy known as DAN, which coerced the chatbot into providing information on illegal activities. Schmidt expressed concern over the lack of a robust non-proliferation framework to mitigate the risks posed by AI. Despite his stark warnings, Schmidt maintained a positive outlook on the broader potential of AI, arguing that the technology is not receiving the recognition it warrants. He reflected on his previous collaborations with Henry Kissinger, noting that they agreed on the monumental implications of AI, which has the potential to surpass human capabilities. He pointed out the unprecedented growth of the GPT series, particularly ChatGPT, which amassed 100 million users in just two months, signaling the immense power of this technology. While concerns about an AI bubble circulate as investments surge in the field, Schmidt believes that history will not repeat itself as it did during the dot-com era. "I don’t think that’s going to happen here, but I’m not a professional investor," he remarked, recognizing the significant economic potential that investors see in AI technology.