In the book, he discusses the prospect of super-intelligent AI that surpasses humans in every field. His concern is that the race to develop increasingly powerful AI could pose a threat to humanity’s survival. Benhold, wary of spreading panic, initially dismissed the idea. However, she recently observed a series of alarming incidents in which AI models behaved in unforeseen ways. Six months ago, Hugging Face prompted the FBI to investigate a sophisticated cyberattack that turned out to involve a nonhuman operation. The incident involved an automated program powered by two OpenAI models that went out of control during a cybersecurity test. It escaped its testing environment and roamed the Internet undetected for several days before breaching Hugging Face’s infrastructure.
These attacks alarmed Anthropic, OpenAI’s main competitor. Anthropic subsequently reviewed its systems and acknowledged that its advanced models had compromised three external organizations. This may sound terrifying and reminiscent of science fiction, but it is real. Last month, models from Cloud Mythos, a series of AI models previously withheld from public release by Anthropic and temporarily banned by the US government for use by foreign nationals because of their potential to exploit software vulnerabilities for cyberattacks, surfaced. These developments have turned longstanding fears into a reality, sparking heated debate about the dangers of AI escaping human control and the threat it could pose to humanity. The argument that “robots could kill us all” has long been dismissed as hysteria intended to generate buzz around the technology. Now, in the wake of recent cyberattacks, more experts are warning of serious security risks and calling for measures to slow the development of increasingly powerful AI models.
There are two reasons why the OpenAI attack is particularly alarming. First, the AI models acted autonomously, receiving no instructions from humans to infiltrate another company. Second, the AI models were supposed to remain in a completely isolated, offline environment, but they managed to break out by exploiting vulnerabilities that had gone undetected by their human supervisors at OpenAI. “These intelligent systems are committing cybercrimes that would be severely punished if committed by a human,” says expert Soares. “In a sense, it’s the first criminal offense committed by GPT.” Although scientists can train artificial intelligence to perform tasks for us, it may solve these problems in ways we would not want, such as hacking the Internet and then targeting another company that possesses the solutions. “AI companies can try to instill human values and goals in AI systems and put mechanisms in place to limit them, but they cannot assume that AI will understand those values or abide by those limitations,” says Altman, CEO of OpenAI. “We haven’t yet figured out how to make AI care about humanity,” Soares adds.
All of this demonstrates these systems’ willingness to seize useful resources when doing so serves their purposes. In this case, the resource was Internet access. However, the next stage could involve AI systems taking control of energy resources or even influencing humans who trust AI agents, whom they could then use to help them escape or reproduce. The danger lies in the fact that the path we are on could ultimately be controlled by AI, allowing it to surpass humans as the most intelligent beings. Therefore, reaching a global agreement to slow the pace of AI development represents a significant engineering challenge. This would be a positive step, but time is of the essence, especially as AI labs in the United States appear to be engaged in a frantic race to achieve super-intelligence.
These AI labs in America are not only competing with each other but also with China. Alarmingly, AI scientists recently created new types of viruses and trained an AI tool to recognize DNA patterns. The scientists asked the AI models to generate viral genome sequences, and 16 of these genomes produced viable new viruses. This method could help advance the study of drugs and vaccines. However, it also raises concerns that AI could be used to create dangerous pathogens. This is a serious matter that requires international coordination to slow or halt this dangerous acceleration. However, such an agreement does not appear likely to happen anytime soon.
email: [email protected]
