Warnings about what artificial intelligence could do in the future — as well as alarming reports about what it has already done — are being issued so quickly that even security experts are struggling to determine what to worry about.
Yet a report released this week by Anthropic’s offensive research team gained attention by describing a practical new threat, one devoid of human characteristics or apocalyptic musings. The Red Team, as these internal testers are called, said it created a malicious AI program that spreads autonomously through internet-connected “smart” TVs and could do the same thing inside other regular devices. While machines running AI typically need a large amount of computing power, the researchers said their program was split into parts among 15 TVs and was still able to run using a little power from each.
In addition to mentioning the horcruxes that kept pieces of Lord Voldemort’s soul alive in objects from the Harry Potter series, the report alarmed some professionals because the generally weak security of smart TVs and many other “Internet of Things” devices would be very difficult to fix, even with a concerted effort.
“This is something to be concerned about, like any other wave of cyberattacks. We need to mitigate the possible effects,” said Joshua Saxe, who led the development of Meta’s AI security efforts. “I am currently concerned that members of the cybersecurity community are not taking these types of real risks as seriously as they should be.”
The damage caused by these “AI worms” could be surprisingly significant. Self-replicating computer viruses, which don’t need to trick a human to spread, are known as worms, and the first worms caused some of the worst havoc in Internet history.
That legacy includes the Morris worm of 1988, which came before the mainstream Internet, when a graduate student released a program that took advantage of various software flaws and weak passwords, copying itself so quickly that many infected machines slowed to a halt. The United States and Israel introduced worms into international espionage with the stealthy Stuxnet in the late 2000s, which sabotaged Iran’s nuclear centrifuges. Other worms, like the Russian NotPetya and the North Korean WannaCry, have caused tens of billions of dollars in damage worldwide.
Many of the world’s most destructive worms depend on known vulnerabilities that device owners have not patched, sometimes months after patches become available. The Anthropic team predicted that a worm like theirs could work the same way and be harder to stop, because smart devices receive patches less often than phones or laptops. Worse yet, they speculated that an AI worm could continue to adapt to other types of hardware by scanning for newly revealed vulnerabilities.
What makes AI worms different from typical self-replicating viruses is that AI models can learn as they go, said Ariel Herbert-Voss, an early security expert at OpenAI who now runs AI security startup RunSybil. Existing cybersecurity infrastructure is not prepared to deal with the proliferation of such threats, Herbert-Voss said. “We need to better empower our security teams. We need to give them more resources to resolve problems.”
Worms could also be used for benevolent purposes, suggested Chris Inglis, former deputy director of the National Security Agency and former national cybersecurity director in the Biden White House.
“The question of whether AI worms are plausible has been settled by several landmark events over the past six months,” Inglis said via digital message. “The better question is how we will use them, counter them and/or combine them with other emerging technologies. »
Anthropic researchers said a successful AI worm should have more capabilities on less powerful devices than is currently possible – but they added that current trends will make this feasible by the end of the year. Their research follows a paper published a few months ago by experts at the University of Toronto, which infected a full-sized computer and could spread to similar machines faster than the TV worm.
“The capabilities developed very quickly,” said co-author Jonas Guan, a doctoral student in Toronto.
after newsletter promotion
Guan and other experts said that, from a practical standpoint, an AI worm should not be high on the list of AI fears. The anthropogenic version moves very slowly, Guan said, and the worms have predictable, detectable actions that can be blocked at multiple levels. Experts say basic security hygiene, including requiring strong passwords and two-factor authentication, would stop most worms and other hacks.
“Nothing about this is really actionable or impactful,” said Juan Andres Guerrero-Saade, vice president of research at security company SentinelOne. “It’s a clever hack that’s far more amplified in the service of ridiculous, doom-laden narratives that suggest the models have a mind and will of their own.”
Anthropic did not respond to an interview request.
Outside of Washington, many in the AI and security world say greater government oversight of AI is now essential. But members of Congress have introduced conflicting bills, while the White House has so far created only a small review committee.
Security researchers and laypeople alike find themselves weighing the motivations behind AI companies’ frightening prognostications, which range from adopting quasi-religious beliefs to seeking legislation to exclude rivals or save money to exaggerating the expected multitrillion-dollar public valuations of San Francisco AI giants.
“The claim that this is largely marketing is correct, but the reality overshadows any public relations,” said Gadi Evron, who is leading an AI security conference in San Francisco this month called Unprompted. “Whatever the labs claim, they don’t give us enough warning. We need to understand that no one can stop the AI storm.”
Gn bussni
