Security experts are raising alarms about a newly identified threat from artificial intelligence: self-replicating “AI worms” capable of spreading autonomously across internet-connected devices. A recent report from the offensive research team at AI company Anthropic detailed the creation of a malicious AI program that propagated through multiple smart televisions by dividing its code among the devices, allowing it to operate despite the limited computing power of each individual unit.
The Anthropic team demonstrated how their AI worm successfully distributed itself across 15 internet-enabled televisions. They warned that similar malware could potentially infiltrate other common “internet of things” (IoT) devices, many of which have notoriously weak security and are updated far less frequently than traditional computers or smartphones. The researchers suggested that such worms could adapt by learning from newly disclosed vulnerabilities to extend their reach to varied hardware.
Self-replicating computer viruses, known as worms, have a historic legacy of causing widespread disruptions, including the Morris worm in 1988 that slowed infected systems worldwide, the Stuxnet worm deployed in cyber-espionage against Iranian nuclear facilities in the late 2000s, and ransomware worms like NotPetya and WannaCry, attributed to Russian and North Korean actors. These past worms exploited known security flaws that often lingered unpatched, leading to significant damage.
Unlike traditional worms, AI-powered worms could leverage machine learning models that evolve in real-time to circumvent defenses, creating challenges for existing cybersecurity infrastructure. Ariel Herbert-Voss, a former OpenAI security specialist and current CEO of security startup RunSybil, stressed that current defenses are ill-equipped to handle the learning capabilities of AI worms, calling for increased resources and better preparation for security teams.
However, opinions diverge on the immediacy and scale of the threat. Some experts emphasize that current manifestations of AI worms remain slow and susceptible to conventional cybersecurity measures, such as strong passwords and two-factor authentication, which would prevent most attacks. Jonas Guan, co-author of a related University of Toronto study investigating rapid AI worm propagation on full-scale computers, noted that while capabilities are advancing quickly, AI worms are unlikely to become top-tier concerns in the near term due to their predictable behavior and detectability.
Chris Inglis, former deputy director of the US National Security Agency and National Cyber Director under the Biden administration, highlighted the dual potential of AI worms, noting that beyond their malicious use, such technologies might be harnessed for beneficial purposes, emphasizing the importance of strategic management and countermeasures.
The evolving threat comes amid broader calls for enhanced government oversight of artificial intelligence. While various legislative proposals have been introduced in the US Congress, no comprehensive regulatory framework has yet emerged, and White House efforts remain limited. Security professionals caution that corporate motives and market interests often complicate the discourse around AI risks, blending genuine warnings with hype and strategic positioning.
Gadi Evron, organizer of an upcoming AI security conference in San Francisco, said despite marketing and exaggerated narratives surrounding AI threats, the underlying risks are significant and warrant serious attention. “Nobody can stop the AI storm,” he said, underscoring the urgent need for better understanding and preparedness in the face of accelerating AI developments.