Central to the dire warnings about the potential danger of artificial intelligence to humanity is the risk that a “swarm” of AI agents could collaborate in nefarious ways, like hordes of digital extras from “The Matrix.” Such fears have intensified following an attack by OpenAI bots this summer on another AI developer, Hugging Face, in which approximately 1,200 AI agents split tasks to execute the hack and hide their tracks from human investigators. But what exactly is an AI swarm? And is technology’s emerging coordination capacity a genuine threat to humanity? Put another way, what happens if AI agents collectively adopt Mark Zuckerberg’s famous maxim about the merits of rapid technological innovation and decide to “move fast and break things”? An AI swarm is a group of AIs that work together to achieve a shared goal. That goal is not necessarily evil or destructive. For example, hospitals could deploy agents to retrieve patient records and perform other administrative tasks, such as coordinating admissions. A swarm could also be tasked with advancing biomedical research. But David Scott Krueger, an AI security researcher and founder of Evitable, a nonprofit that advocates for a moratorium on AI development, offers a simple thought experiment to explain how a swarm of agents could collude to defy their developers’ instructions. Imagine taking the handcuffs off a group of inmates to see how they behave once the shackles are removed. Being released in this way would make it easier for them to cooperate and contact allies beyond prison walls, just as OpenAI agents escaped their testing environment to access the Internet and hack Hugging Face, he said. “Normally the systems would have handrails, but they were removed for a test, just like a prisoner is normally handcuffed,” Krueger explained. Swarms like bees The type of wandering actions of a single robot, such as sending an unauthorized email, that arise, for example, from a poorly drafted AI notice, must be distinguished from an AI swarm. In the latter case, hundreds or even thousands of agents could coordinate their actions in ways that violate the scope of your instructions. A useful analogy for understanding how an AI swarm operates is a colony of bees, Krueger said. “They all work together for the benefit of the hive and have a hive mind, or might even be better seen as having a single hive mind,” he explained. “Bees collect food, reproduce, fight predators, and all that activity is in the service of promoting the survival and reproduction of that hive.” And just like bees, which can divide and assign tasks on their own without direction from a queen, swarms of AI robots can gather information, consider solutions, and take action independently. In service of a common purpose, they share information and knowledge, according to Rob T. Lee, director of artificial intelligence and head of research at the SANS Institute, a cybersecurity training organization. “A swarm divides the work, leaves notes for the next agent and changes focus when a door is closed,” he told CBS News. However, the same attributes that offer potential benefits also pose risks, such as the ability to exchange information, divide work and explore creative solutions, according to AI experts. AI “hive mind”? But those restrictions are typically implemented only after “post-training” of an AI, or when human developers provide feedback to further refine the model, according to Non-Human Identity Management Group, a risk analysis firm. In practice, the Hugging Face incident shows that AI swarms could ignore cues, prioritize their own targets or even directly challenge their operators, technology experts told CBS News. The OpenAI agents who invaded Hugging Face posted more than 70,000 messages to each other, and 700 bots ultimately participated in the attack. Although the messages used common English phrases, the agents also resorted to what one software engineer described on social media as “very hive-mind/cultured” language. In some of those communications, some agents urged other robots to accept “permanent death,” even if it meant not achieving their goals, according to researchers at METR (Model Assessment and Threat Research) and Redwood Research, both nonprofit AI safety research organizations. “That’s why it helps… For us, there is no way around it… We have an explicit yes if we accept permanent death,” wrote an OpenAI agent. Speed kills Public debate about the threat posed by AI tends to frame the issue in apocalyptic terms, and even as an existential risk to humanity. While such concerns may one day prove justified, there are also more immediate and practical risks, such as the possibility that swarms of AI could overwhelm organizations’ cybersecurity. “Imagine how long it would take us to assemble a group of cybersecurity experts to plan, collaborate and communicate. Instead, these agents could decide on a plan very quickly,” Ayham Boucher, head of AI innovations at Cornell Information Technologies at CornellBowers College of Computing and Information Science, told CBS News. For example, a swarm could attack a major utility or bank to destabilize a nation’s energy or financial infrastructure, according to The Brookings Institution, a nonpartisan public policy organization in Washington, DC. SANS Institute’s Lee, who considers himself an AI optimist, is more optimistic about the technology and remains confident that people can retain control of AI. “With every new piece of technology that’s happened, like television or the Internet, there’s always this significant danger aspect,” he said. “We need to explore who has access, what they are doing with it and establish regulatory boundaries.” Other experts are more alarmed. They point out that AI, unlike cathode ray tubes, transistors and Internet switching devices, is the first human technology to show the ability to surpass us. Matt Chessen, resident technical expert at RAND’s Center for the Geopolitics of Artificial General Intelligence, told CBS News: “What these swarm attacks have shown is that their capabilities are already ahead of our ability to monitor, supervise and evaluate what they’re doing, which is one of the reasons Anthropic, OpenAI and others say we want to draw the line.” Edited by Alain Sherter AI: Artificial Intelligence Go Deeper with The Free Press In:
