In a novel cybersecurity experiment, researchers at the artificial intelligence company Anthropic created a malicious AI program that spread across 15 internet-connected smart televisions. Because advanced AI models typically require substantial computing power, the company's internal offensive testing team—known as a red team—divided the program into separate pieces. By drawing small amounts of processing power from each television, the distributed worm was still able to function and replicate autonomously.
Unlike traditional computer worms that spread by exploiting fixed software flaws, an AI-powered worm could theoretically adapt as it moves by researching newly disclosed vulnerabilities. Security specialists are particularly concerned about internet-of-things devices, such as smart televisions and household appliances, because they receive security patches far less frequently than smartphones or laptops. Joshua Saxe, who previously led Meta's AI security development, warned that the cybersecurity community is not taking the real risks of such self-propagating threats seriously enough.
However, independent security analysts disagree sharply over whether such attacks pose an immediate practical danger. Jonas Guan, a doctoral researcher at the University of Toronto who co-authored earlier work on AI worms, noted that Anthropic's test program moves very slowly and carries out predictable actions that standard network defenses can intercept. Juan Andres Guerrero-Saade, vice president of research at security firm SentinelOne, dismissed the project as "nifty stunt hacking" meant to amplify dramatic narratives about autonomous software, arguing that basic digital hygiene such as strong passwords and two-factor authentication can block most worms.
Anthropic researchers acknowledged that running capable AI worms on low-power consumer devices remains difficult today, though they argued that rapid improvements in AI efficiency could make the technique feasible in the near future. While former US National Cyber Director Chris Inglis suggested that AI worms have become technically plausible and could eventually serve defensive as well as offensive purposes, the test has intensified debates over whether governments must introduce stricter oversight before automated threats mature.