Anthropic Red Team Demonstrates Self-Replicating AI Worm on Smart TVs
Anthropic researchers demonstrated an experimental AI worm that splits itself across 15 smart TVs, sparking industry debate over emerging threats to connected devices.
A research demonstration from Anthropic's offensive security group has brought renewed attention to the potential risks of self-replicating artificial intelligence. The company's red team successfully built an experimental AI program that traversed internet-connected smart televisions without human intervention. Instead of running entirely on a single machine, the program fragmented parts of itself across 15 smart TVs, drawing small amounts of computing capacity from each device to remain functional.
The demonstration highlights persistent vulnerabilities within the Internet of Things (IoT). Consumer hardware such as smart televisions often features minimal security safeguards and receives software patches far less frequently than personal computers or mobile phones. While AI systems traditionally demand substantial computational resources, Anthropic's team warned that technical improvements could allow similar self-propagating software to operate effectively on low-power devices by the end of the year.
The Mechanics and Evolution of AI Worms
Self-propagating code has historically produced some of the most disruptive cybersecurity incidents. Precedents range from the 1988 Morris worm, which overwhelmed early networked systems, to the Stuxnet worm used against Iranian nuclear centrifuges in the late 2000s, as well as destructive outbreaks like WannaCry and NotPetya. These attacks succeeded by exploiting unpatched flaws across connected machines.
Anthropic's researchers suggested an AI-driven worm could adopt similar pathways while dynamically analyzing newly disclosed software vulnerabilities to hop across disparate hardware environments. Ariel Herbert-Voss, an early OpenAI security specialist who now leads the AI security company RunSybil, noted that AI models can adapt and learn over time, separating them from static legacy malware. Herbert-Voss argued that existing defensive setups remain under-resourced to address such evolving vectors.
The Anthropic trial follows earlier academic work from researchers at the University of Toronto, who previously demonstrated an automated worm spreading across standard computers faster than the television prototype. Co-author Jonas Guan, a doctoral student at Toronto, stated that relevant technical capabilities are progressing rapidly, though he noted the Anthropic demonstration currently operates at a very slow pace.
Industry Debate: Practical Threat vs. Stunt Hacking
The research has sparked differing perspectives within the cybersecurity community. Former White House National Cyber Director Chris Inglis observed that breakout incidents over the past six months have proven the plausibility of AI worms, suggesting future debates will center on how they are utilized, countered, or integrated with other technologies.
Similarly, Joshua Saxe, former head of AI security at Meta, urged the security community to treat these developments with urgency, emphasizing the need to mitigate potential risks before they mature. Gadi Evron, organizer of the Unprompted AI security conference, also maintained that the underlying technical reality outweighs public relations hype.
Conversely, some industry veterans remain unconvinced of the immediate danger. Juan Andres Guerrero-Saade, vice-president of research at SentinelOne, characterized the demonstration as "nifty stunt hacking" that amplifies alarmist narratives rather than reflecting a viable real-world attack vector. Analysts also noted that AI worms execute recognizable and predictable patterns that standard defensive measures can counter. Routine security hygiene, including regular patch deployment, strong passwords, and multi-factor authentication, remains sufficient to halt most replication attempts.
What it means for developers
For software engineers and system architects, the demonstration serves as a reminder of the fragility of edge devices and distributed networks. Connected hardware lacking robust authentication and regular update cycles presents an accessible target for automated exploitation, particularly as model footprints decrease and local inference capabilities improve.
Developers building connected services must prioritize foundational security practices. Restricting automated lateral network movement, enforcing multi-factor authentication, and ensuring timely firmware updates remain the most effective defenses against autonomous propagation. As engineering teams evaluate defensive architectures and build applications leveraging modern language systems, developers can try top AI models cheaply through one API at https://apixoai.online to assess how current models handle structured security tasks and vulnerability analysis.
Ultimately, while experts agree an automated IoT worm is unlikely to cause immediate widespread disruption due to current execution speed and existing countermeasures, the Anthropic experiment shows that distributed compute models will continue to influence both offensive and defensive software design.
Source: What are self-replicating ‘AI worms’, and why is Anthropic worried about them? — The Guardian AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
