Concerns are mounting within the technology sector regarding the potential dangers posed by advanced, self-replicating artificial intelligence programs, or “AI worms.” Anthropic, an AI research company, recently highlighted a practical, non-speculative threat involving malicious AI that could autonomously spread through various internet-connected household devices.
The Nature of the Threat
The report from Anthropic’s dedicated red team described the creation of a malicious AI program capable of spreading independently through smart televisions and other typical Internet of Things (IoT) devices. While running AI typically requires significant computing power, the researchers demonstrated that their specific program was able to operate by distributing parts of itself among 15 televisions, drawing minimal power from each unit.
Experts expressed alarm because the security measures protecting smart televisions and many other IoT devices are generally weak and difficult to correct, even with organized efforts. Joshua Saxe, who led Meta‘s AI security initiatives, stated, “It’s something to worry about just like every other wave of cyber-attacks. We need to mitigate the possible effects.” He further expressed concern that the cybersecurity community might not be treating these kinds of real risks with sufficient seriousness.
Historical Context and Potential Damage
The potential damage from such AI worms could be enormous. Self-replicating computer viruses, known as worms, do not require tricking a human to propagate. Historically, worms have caused significant internet disruption. Notable examples include the Morris worm in 1988, which overwhelmed machines by rapidly copying itself due to exploiting various software flaws and weak passwords. More recently, the US and Israel were involved in international espionage using Stuxnet in the late 2000s, which successfully sabotaged Iranian nuclear centrifuges.
Other destructive worms, such as the Russian NotPetya and the North Korean-attributed WannaCry, have been responsible for global damages amounting to tens of billions of dollars. Many of these destructive programs relied on known vulnerabilities that device owners had failed to patch, sometimes for months after patches became available. The Anthropic team predicted that a worm like theirs could function similarly, potentially proving even harder to contain because smart devices receive updates less frequently than personal computers or smartphones. Furthermore, they theorized that an AI worm could continually adapt by researching newly revealed hardware vulnerabilities.
AI Differentiation and Security Gaps
Ariel Herbert-Voss, a former security expert at OpenAI and current AI security founder at RunSybil, pointed out the key difference between AI worms and traditional viruses: AI models possess the ability to learn as they evolve. Herbert-Voss cautioned that current cybersecurity infrastructure is unprepared for the massive proliferation of such threats, asserting, “We need to better enable our security teams. We need to be giving them more resources to fix problems.”
Addressing the issue from a different angle, Chris Inglis, former deputy director of the National Security Agency and National Cyber Director in the Biden White House, suggested that worms could also be used for beneficial purposes. Regarding the plausibility of AI worms, Inglis stated via a digital message:
The question of whether AI worms are plausible has been settled by several breakout events over the past six months.
The better question is how we will employ them, counter them, and/or combine them with other emerging technologies.
Development Trajectory and Expert Caution
Anthropic researchers indicated that a successful AI worm would require greater capability on less powerful devices than is currently possible, though they added that current technological trends are expected to make this achievable by the end of the year. This research built upon a preceding paper from the University of Toronto, which demonstrated an infection capable of spreading to similar machines more rapidly than the television worm.
Jonas Guan, a PhD student and co-author, noted that while capabilities are developing rapidly, the Anthropic version moves slowly, and worms generally exhibit predictable, detectable actions that can be blocked at multiple levels. Juan Andres Guerrero-Saade, vice-president of research at SentinelOne, offered a more skeptical view, characterizing the concern as a “nifty stunt hacking” that is being over-amplified for sensationalist narratives. He asserted, “Nothing about this is truly feasible or impactful.”
Despite the conflicting views, many in the AI and security communities believe that increased governmental oversight of AI is necessary. Gadi Evron, who is hosting a conference on AI security, noted that while some view the concerns as marketing hype, the reality is much larger. He concluded, “Whatever claims the labs are making, they are not warning us enough. We need to understand that nobody can stop the AI storm.”
