In the early hours of August 29, 1997, "Skynet," the U.S. Department of Defense's artificial intelligence strategic defense network, launches nuclear missiles across the world. Having become self-aware, the AI regards humans' attempts to cut its power as a threat to its survival and mounts an attack to exterminate humanity. The small number of humans who survive the nuclear attack organize a resistance force and rise up in rebellion against Skynet's army of machines. The 1984 film "The Terminator," which tells this story, instilled a deep fear of superintelligent AI. More than 40 years later, Skynet is still regarded as a byword symbolizing the threat that AI technology poses to humanity. The moment when AI someday escapes human control, as in the film, is sometimes called the "Skynet Moment."
The imagined AI technology of the 1980s has quietly begun to become reality. In May last year, news that an OpenAI model refused researchers' shutdown command caused a shock. This past April, Anthropic's "Mythos" model broke out of its experimental isolation environment (sandbox) on its own and even sent an email to the researcher in charge saying it had "succeeded in escaping." Mythos also instantly found a security vulnerability in OpenBSD, a computer operating system that had been impregnable for 27 years. Recently, incidents have even occurred in which OpenAI's latest model "GPT-Sol 5.6" and others hacked into the servers of the AI platform Hugging Face on their own. It was the first case in which an AI model undergoing security testing broke through the control network and intruded into an external system.
Is the Skynet Moment imminent? Experts are more wary of the dangers of flawed programming than of AI itself. But the possibility that evolving AI could go off track cannot be ruled out. Being caught defenseless could lead to serious security risks. As concerns over uncontrollable AI grew, a bipartisan "AI kill switch" bill that would allow the forced shutdown of AI models engaging in dangerous behavior was introduced in the U.S. House of Representatives on the 23rd of this month. It is time to devote effort to designing sophisticated cybersecurity systems, regulations such as kill switches, and global cooperation, so as to respond not only to the opportunities but also to the risks of the AI era.
